Biochips For Microfludics

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Microprocessors and Microsystems 98 (2023) 104774

Contents lists available at ScienceDirect

Microprocessors and Microsystems


journal homepage: www.elsevier.com/locate/micpro

A bidirectional droplet routing in digital microfluidics biochip


Jyotiranjan Swain ∗, Sumanta Pyne
Department of Computer Science and Engineering, National Institute of Technology, Rourkela, Odisha, India

ARTICLE INFO ABSTRACT

Keywords: Droplet routing in digital microfluidic biochip is very challenging. It involves moving droplets from its source
Droplet routing cell to destination cell. The challenge is to maintain fluidic constraints for every droplet movement. In this
Bidirectional article, we have proposed a greedy bidirectional droplet routing based on divide and conquerer approach. The
Bipartite graph
biochip is explored using two waves i.e. one from source cell and other from destination cell. Then, the biochip
Pivot cell
is converted into a bipartite graph by identifying strongly connected components. The route selection is done
Strongly connected component and greedy
selection
using our four greedy selection heuristics. The final route is generated by appending multiple sub-routes. The
simulation result show an improvement of 7.32% and 18.05% in latest arrival time and number of shared cells
respectively in free topology. Similarly we register an enhancement of 8.43% and 19.70% latest arrival time
and number of shared cells respectively in virtual topology.

1. Introduction appending sub-routes. Simulation result shows significant improvement


in latest arrival time and number of shared cells.
Digital microfluidic biochip (DMFB) is one of trending research area This article made following contributions
now a days. DMFB is a miniature hand-held device, that automate
• We use bidirectional wave propagation for exploring the whole
biochemistry assay synthesis [1,2]. DMFB consists of two glass plates
biochip.
held on top of each other. The fluid droplets are sandwiched in the gap
• Strongly connected components are identified and biochip is
between the plates. The surface of the glass plates are painted with
transformed into a bipartite graph model.
hydrophobic paint. This is done to avoid sticking of residual fluid from
• A new heuristics 𝑃 𝑖𝑣𝑜𝑡𝑉 𝑎𝑙𝑢𝑒 is defined for route selection.
a droplet on the glass plate. The top-view and side-view of the biochip
• We store the whole explored graph after route identification.
is presented in Fig. 1.
• When deadlock occurs, the stored sub-routes helps to find routes
DMFB constitutes a set of electrodes (called cells) arranged in two
for faster detouring.
dimensional array manner. Individual droplets reside on top of a cell
and can only move to a non-diagonal adjacent cell using electro-wetting This article is organized as follows. Section 2 introduces droplet routing
principle [3,4]. Fig. 2 presents various stages of a biochemical synthesis and its challenges in details. Section 3 lists literature survey of previous
using DMFB. The stages are scheduling, placement and routing. A bio- works. Section 4 elaborates our proposed method in depth. Simula-
chemical assay is represented using directed acyclic graph (DAG). In the tion, experimental result and performance analysis given in Section 5.
DAG, nodes represent the fluidic operations such as mix, dilute, split, Finally, conclusion and future work is presented in Section 6.
move, etc. The edges represent the interdependency among various
operations [5–7]. In scheduling phase, all assay operations are assigned 2. Droplet routing
starting and ending time cycle. During placement phase, modules are
arranged in optimal manner, so that the droplet traveling distance Droplet routing in digital microfluidic biochip is analogous to tra-
is minimized. In routing phase, biochemical assays are executed by ditional computer network routing. In both cases, the goal is to move
physically moving various droplets. In this phase, the routes of in- the fluid droplet or data packet from their source cell or node to the
dividual fluid droplets are explored, then droplets are moved from destination cell or node. Few notable differences between traditional
their source cell to their target cell along their route [3,8–10]. In this computer network routing and droplet routing is given below
paper we proposed a greedy bi-directional routing method using wave
propagation for route exploration. The routes are then selected by using • DMFB is a rectilinear grid while traditional computer network can
one of four user defined heuristics. The final route is generated by have any topology.

∗ Corresponding author.
E-mail address: jrswain85@gmail.com (J. Swain).

https://doi.org/10.1016/j.micpro.2023.104774
Received 28 August 2021; Received in revised form 21 July 2022; Accepted 8 November 2022
Available online 4 February 2023
0141-9331/© 2023 Elsevier B.V. All rights reserved.
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Fig. 1. Structure of digital microfluidic biochip. (a) Side-view, (b) Top-view.

Fig. 2. Biochemical synthesis stages using DMFB.

Fig. 3. Complexity in droplet routing. (a) Multiple routes having same length for a droplet, (b) Multiple droplets each having multiple routes.

• In DMFB, only rectilinear movements are allowed i.e fluid


List of Symbols
droplets can only move to top, bottom, left or right cells and
no diagonal movement is allowed while in traditional computer 𝐷𝑀𝐹 𝐵 Digital microfluidic biochip
network all movements are allowed. 𝐷𝐴𝐺 Directed acyclic graph
• If two fluid droplets occupy two adjacent cells, they will mix with 𝐹 𝐷𝑖 𝑖th droplet
each other. This unwanted mixing may lead to erroneous result.
𝑆𝑖 Source cell of 𝑖th droplet
So in order to avoid this, two fluid droplets can’t occupy two
𝐷𝑖 Destination cell of 𝑖th droplet
adjacent cells and a space of single cell has to be maintained be-
𝑅𝑖 Route of 𝑖th droplet
tween two droplets. This is property is called fluidic constraint. In
traditional computer routing, two packets can be placed adjacent 𝐿𝐴𝑇 Latest arrival time
to each other.

The routes in DMFB are explored by using Manhattan distance


instead of euclidean distance. For a given source–destination pair, there can be represented only by storing its two endpoints. Similarly, a route
may exist multiple routes [2,11]. Only one among these routes need to having a single bend can be represented by three entries i.e. start point,
be selected. Route selection is influenced by various parameters such bending point and end point. When the objective is to have optimal
as route length and traversal time. In this article, we have proposed memory, then a route with least number of bends is preferred [12].
two new heuristics i.e. number of bends and number of intersecting points. The second heuristic number of intersecting points specifies the number
The metric number of bends influences the memory requirement to store of shared cells in a route. If the number of shared cells is high, then
routes. A bend occurs in a route, when there a change in direction. For it lead to contamination problem [13] and vice-versa. To avoid the
example, a droplet is traversing cells along X axis and moves to a cell contamination problem, routes with minimal number of intersecting
in Y axis or vice versa then a bend occurs. A route having no bends, points is preferred. As the number of droplets increases, the complexity

2
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

These solutions are not applicable for many scenarios and produce
optimal solution in some cases. The most popular domain in proposed
solutions belongs to graph theory. Many popular routing solutions are
demonstrated in [4,12,17–19,19–25]. Many Swarm intelligence based
works are also listed in [9,26].
VLSI routing based algorithm able to run small assays(PCR,In-vitro)
but failed for large assays (Protein split). The unavailability of free
cells for larger assays is behind this shortcoming. Network model based
methods suffers from bottleneck of placement of modules, droplet in-
jection rate or storage issues, that leads to the complexity in converting
routing problem in flow models. Swarm intelligence based works able
to run all benchmarks but their execution time is very high. So, these
protocols applicable to off-line system. Graph based solutions gives
optimum value for on-line as well as off-line systems. Few routers
like Roy routers, Grissom Routers, Huang router [16] able to run all
benchmarks. Grissom router is designed for virtual topology while
Swain router [12] is tailored for free topology. There are two types
of DMFB systems such as online and offline. In online system the
routes are explore dynamically while the offline systems the routes are
predefined. Online systems have a deadline to complete an operation
while offline systems don’t have such constraint. So offline systems
prefers to have minimal length routes, while online systems prefer less
Fig. 4. Overview of proposed algorithm.
congested route. In this work, our goal is to design an algorithm which
will have optimal value for both on-line and off-line systems.

4. Proposed algorithm
of selecting routes for individual droplets becomes higher. To illustrates
this, let us consider an example presented in Fig. 3(a). We have a single
Disjoint routes plays an important role in droplet routing. A sched-
droplet with source cell S and destination cell D. The task is to find a
ule consisting of only disjoint routes is highly desired. In this schedule,
route between cells S and D. The shortest route to reach from S to D
the occurrence of deadlock becomes zero. The probability of finding
have length 7. There are many such routes, only three routes 𝑅1 , 𝑅2
disjoints routes reduces drastically as the bio-chemical assay density
and 𝑅3 given in Fig. 1(a). number of bends in the routes 𝑅1 , 𝑅2 and 𝑅3
increases. Practically, most schedules contains mostly shared routes. In
are 1, 5 and 3 respectively. If the goal is to have lowest runtime then
an optimal route map, the number of shared cells should be minimized.
any one among the three route can be selected. If the objective is to
Our proposed algorithm’s goal is to optimize the number of shared cells.
reduce memory requirement, then route 𝑅1 is preferred. As the number This results in minimal latest arrival time for an assay.
of droplets increases, the complexity in the selection of routes increases.
In Fig. 1(b), there are three droplets or source and destination pair and 4.1. Overview of proposed algorithm
seven routes. The final route map will have only three routes i.e. single
route for every source and destination pair. In final route map have In scheduling phase, every operation in the DAG is assigned both
two intersecting routes and a disjoint route. If we have 100 droplets start and end time cycles/periods. To execute various operations such
and 500 routes, we need to reduce the route map to have 100 routes as mixing, splitting, heating specific modules are assigned. In placement
(single route for each droplet). The selection of these 100 routes need to phase, the modules are arranged in such a way that the travelling time
be done in such a way that the number of bends and intersection points for droplets are minimized. After scheduling and placement phase, we
should be optimized. The droplet routing problem is now converted have a set of droplets and operations. Each operation has beginning as
into a route optimization problem. well as ending time period the module where the operation is to be
performed. This information is given as input to the droplet routing
2.1. Problem formulation phase. The overview of our proposed method is illustrated in Fig. 4.
Our proposed algorithm have three phases i.e. route exploration,
Mathematically droplet routing problem can be defined as route selection and route compaction. The route exploration phase use
Input: A m×n biochip with B blockages, a set of k droplets FD. Each wave propagation method to explore the whole biochip. Our proposed
droplet F𝐷𝑖 , have a source cell (𝑋𝑖𝑠 , 𝑋𝑖𝑠 ) and a destination cell (𝑋𝑖𝑑 , 𝑌𝑖𝑑 ) method uses two waves for route exploration. The first wave starts
where i ≤ k. from source cell and move towards the destination cell. The second
Output: A set of route R containing a route 𝑟𝑖 of every droplet F𝐷𝑖 . wave originate from destination cell and proceed towards the source
Constraint : The fluidic constraint has to be maintained at every cell. Both the waves are propagated in round robin fashion. In an
instant of time. iteration, both the waves propagate by a single cell. Every cell visited
is marked by an alpha numeric value i.e. 𝑋𝑎. Where 𝑋 is either S
3. Related works or D specify source or destination wave, and 𝑎 is a numeric value to
designate the iteration number. The exploration phase is terminated,
Bohringer introduces an algorithm based on 𝐴∗ algorithm for when all the branches of propagating wave encounter a cell that is
droplet routing. Su et al. [1] then demonstrated the fluidic constrains in marked 𝑋𝑎 or no free cells are available. The cells marked with both
droplet routing physically. Many authors proposed droplet routing pro- 𝑆𝑎 and 𝐷𝑏 are then identified. These cells are called 𝑝𝑖𝑣𝑜𝑡𝑐𝑒𝑙𝑙𝑠. The
tocols for digital microfluidic biochip. These works can be grouped in pivot cells represent the intersection points of both propagating waves.
to four groups namely VLSI, Network Flow, Graph and Miscellaneous. These cells act as a boundary between cells marked by source wave and
Protocols based on VLSI routing are presented in [1,14,15]. Moving destination wave. This transforms the rectilinear grid to a bipartite grid.
droplets on the biochip surface emulates network flow model. Works Any route connecting source cell S and destination cell D must have
illustrated in [2,3,5,6,8,16] propose solutions based on this model. one of the pivot cells. This make the pivot cells as a strongly connected

3
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Fig. 5. This example demonstrates wave propagation in route exploration process. (a) A 10 × 8 Biochip with five droplets (each pair is uniquely colored), (b)–(f) Waveform
propagation of FD1 , FD2 , FD3 , FD4 and FD5 respectively.

components. We defined a new heuristic called 𝑝𝑖𝑣𝑜𝑡𝑣𝑎𝑙𝑢𝑒 to generate (called ants) are generated from the selected pivot cell. One ant travel
a priority order among pivot cells. 𝑝𝑖𝑣𝑜𝑡𝑣𝑎𝑙𝑢𝑒 is calculated by summing towards source and the other proceed towards destination. Each ant
both the iteration values in a pivot cell i.e. 𝑆𝑎 and 𝐷𝑏 . The pivot values move to cell that have value minimum iteration value. Both these ants
are then sorted in ascending order. This generates a local ordering are terminated when they reached source or destination cell. Each ant
among the droplets. Pivot cell(s) with minimum pivot value(s) have the moves one cell at a time and append the traversed cell address in its
highest priority. Each droplet may have multiple routes having same header. When one of the ant reaches the source or destination, the sub-
pivot value. We have proposed four new greedy heuristics for route route is extracted from header and stored at the source or destination
selection. In greedy approach, we select one or multiple pivot cells cell. When both ants reaches their endpoint, final route is generated
based on the selection heuristic. In route tracing phase, two packets by appending both the sub-routes. After selection of a single route for

4
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Fig. 6. This example illustrates final result of route selection. (a) Total random (b) Single greedy (c) Multiple greedy (d) All greedy.

Table 1
Latest arrival time in free toplology.
Benchmark # Nodes Pan2011[9] Grissom2018 [24] Juarez2018[26] Swain2019 [12] Huang2020 [16] Total random Single greedy Multiple greedy All greedy
PCR 16 785 517 839 562 604 516 638 795 834
In-Vitro 1 20 863 572 860 581 627 628 645 817 892
In-Vitro 2 30 709 624 917 613 652 687 683 854 926
In-Vitro 3 45 956 893 1425 874 929 1005 1120 1210 1381
In-Vitro 4 60 1521 1380 1906 1469 1382 1382 1284 1356 1446
In-Vitro 5 80 1954 1632 2944 1935 2012 2067 1687 1631 1872
Protein Split 1 28 726 641 832 638 649 659 684 709 951
Protein Split 2 58 1673 1426 1519 1724 1626 1461 1464 1585 1907
Protein Split 3 118 8450 7809 8126 7983 8064 7810 7809 8236 9086
Protein Split 4 238 19736 Fail 18924 17657 18731 18975 17472 18124 19258
Protein Split 5 478 Fail Fail Fail 38246 41869 40219 37987 37657 38463
Protein Split 6 957 Fail Fail Fail 75591 Fail 77621 74487 73549 71876
Protein Split 7 1938 Fail Fail Fail 202317 Fail 253420 199287 201637 194531

every droplet, the route map is generated. Finally, route compaction is destination. Waves are propagated one cell in a time cycle in a round
performed to generate a parallel moving sequence. robin fashion. While propagating each cell is marked by 𝑆𝑖 or 𝐷𝑗 where
𝑖, 𝑗 are iteration number and S is source and D is destination. The wave
4.2. Route exploration propagation of all the droplets is presented in Fig. 5(b–f). After the
wave propagation terminates, we search for cells which are marked
Let us consider the 10 × 8 biochip in Fig. 5(a). The biochip contains both 𝑆 and 𝐷. These cells are called pivot cells and highlighted in
two types of cells i.e. free cell and blockages. A free cell can be occupied Fig. 5(b–c). The set of pivot cells acts as a boundary in grid of biochip.
by a fluid droplet anytime. A blockage cells can’t be occupied by a fluid This transforms the biochip into a bipartite graph. One side contains
droplet anytime. Free cells are shown as white cells and blockages are 𝑆 marked cells and the other side contain 𝐷 marked cells. Every pivot
represented as black cells. There are five droplets i.e. FD𝑖 , 0 < 𝑖 ≤ 5. cell is assigned 𝑝𝑖𝑣𝑜𝑡𝑣𝑎𝑙𝑢𝑒 by adding the marked propagation iteration
Each droplet is represented by same colored source and destination cell values. Then all the droplets are sorted in ascending order using pivot
pairs. In route exploration, each droplet is considered separately. The value. This results a priority order of the droplets. In other words, the
route exploration starts from the highest priority droplet. First, two droplet having least pivot value have highest priority and vice-versa.
waves are generated. One from source cell and other one from the The droplet order for the biochip in Fig. 5(a) is FD3 (7) > FD5 (9) >

5
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Table 2
Latest arrival time in virtual toplology.
Benchmark # Nodes Pan2011[9] Grissom2018 [24] Juarez2018[26] Swain2019 [12] Huang2020 [16] Total random Single greedy Multiple greedy All greedy
PCR 16 1037 724 1156 829 726 771 784 817 869
In-Vitro 1 20 1165 1052 1201 1324 1147 953 951 976 1023
In-Vitro 2 30 1426 1269 1434 1721 1560 1082 1037 1041 1108
In-Vitro 3 45 2518 2274 2873 2769 2303 1973 1826 1871 2075
In-Vitro 4 60 3642 3519 3027 Fail 4302 3901 3917 3960 4102
In-Vitro 5 80 8496 8637 8843 Fail 8076 8301 8309 7645 7967
Protein Split 1 28 1752 1632 1829 2056 1761 1612 1746 1845 1870
Protein Split 2 58 4045 3175 3756 4215 3676 3950 3406 3264 3342
Protein Split 3 118 10297 Fail 9984 Fail 9872 10211 10467 9843 9807
Protein Split 4 238 Fail Fail Fail Fail Fail 23391 22836 21378 21532
Protein Split 5 478 Fail Fail Fail Fail Fail 49103 47621 43276 43174
Protein Split 6 957 Fail Fail Fail Fail Fail 103244 101353 100642 99265
Protein Split 7 1938 Fail Fail Fail Fail Fail Fail 321457 316521 267760

Fig. 7. Latest arrival time in free topology (a)small benchmarks i.e. PCR,In-Vitro 1–5 and Protein split 1–2 (b) Large benchmarks i.e. Protein split 3–7.

6
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Fig. 8. Latest arrival time in virtual topology (a) Small benchmarks i.e. PCR,In-Vitro 1–5 and Protein split 1–2 (b)Large benchmarks i.e. Protein split 3–7.

FD4 (10) > FD2 (11) > FD1 (16) where FD𝑖 is droplet 𝑖 and the integer 4.3.1. Total random
represents pivot value. Here we don’t take pivot value into consideration and select one cell
randomly. This cell can have any value in the pivot value set.

4.3. Route selection 4.3.2. Single greedy


If we have multiple number of minimum value pivot cells then
we choose only one pivot cell randomly. Droplet FD1 have six pivot
The pivot cells act as a strongly connected component. All the final cells and every pivot cell have value 16. Since all cells is a potential
route must have one of the pivot cells. Route selection starts with mini- candidate, one is chosen randomly.
mum pivot value i.e. the highest priority droplet. There may be multiple
droplets having with minimum pivot value. So, the selection of pivot 4.3.3. Multiple greedy
cell is very vital. Here we proposed four ways of selection of pivot cells If we have multiple number of minimum value pivot cells then we
i.e. total random, single greedy, multiple greedy and all greedy manner. choose all the pivot cells with minimum value. The pivot value of

7
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Table 3
Number of shared cells in free topology.
Benchmark Pan2011[9] Grissom2018 [24] Juarez2018[26] Swain2019 [12] Huang2020 [16] Total random Single greedy Multiple greedy All greedy
PCR 436 721 485 317 382 621 570 421 324
In-Vitro 1 505 854 531 346 425 680 586 462 345
In-Vitro 2 672 1063 829 693 675 734 791 670 663
In-Vitro 3 1324 1737 1164 994 1034 1427 1248 986 1021
In-Vitro 4 1657 1856 1386 1570 1657 1943 1341 1328 1386
In-Vitro 5 4260 5037 4219 2486 3503 4461 4260 4363 4209
Protein Split 1 641 511 563 509 521 624 507 651 523
Protein Split 2 1324 1157 1201 1148 1264 1143 1261 1220 1187
Protein Split 3 2359 5241 4860 3365 3479 3321 5033 4154 3275
Protein Split 4 4463 Fail 9745 8470 10536 11054 9765 8501 7546
Protein Split 5 Fail Fail Fail 12304 23021 18760 12801 13644 11905
Protein Split 6 Fail Fail Fail 24221 Fail 32006 30524 26701 23642
Protein Split 7 Fail Fail Fail 49518 Fail 53474 67294 60362 47370

FD2 consists of set { 11, 11, 11, 13, 13, 13, 15, 17}. The pivot cells pivot cell. Route generation is defined in step 20–23. Each ant has a
(5,5), (6,6) and (7,7) have values 11. So these three pivot cells selected sub-route header, used for storing address of each cell it traveled. The
as potential cell in the final route. first ants that reaches the source or destination, its sub-route header is
extracted. Both subroutes are then appended together to generate the
4.3.4. All greedy final route.
Here the pivot value is made irrelevant. So all the pivot cells are Algorithm 1: : Route Exploration and Route Selection
candidates. The pivot value of FD5 consists of set {9, 9, 11, 13, 15,17}. Input : A list of nets, set of free cells and set of blockages.
Here all cells are considered for selection. Output: A set of routes
1 /* Route Exploration */
4.3.5. Route tracing 2 for net 𝑁𝑗 : 𝑗 = 0 to 𝑛 − 1 do
After selection of pivot cell(s), two packets called 𝑎𝑛𝑡s are generated 3 Generate two waveform (one from source cell and one from
from every selected pivot cell. One ant moves towards source and other target cell)
ant towards destination. Every ant moves a cell per time cycle. The next 4 Terminate propagation if free cell is not available.
cell to be selected must have minimum iteration value than the current
5 end for
iteration value. The iteration value acts as pheromone for ants as in
6 for net 𝑁𝑗 : 𝑗 = 0 to 𝑛 − 1 do
ant colony optimization. For every cell traversed by an ant, the cell
7 Find pivot cells and assign pivot value.
address is appended to its header. If ant 𝐴 is at cell 𝐶 and moves to
8 Sort the pivot values in ascending order.
cells 𝐾 through 𝑋&𝑌 cells, the header field will have value {C,X,Y,K}.
9 Find the min Pivot value of every droplet.
The ants terminates when they reach their target. When an ant reaches
10 end for
its target, the route is extracted from its header and stored at source
and destination. Final route is generated by combining both sub-routes
11 /* Droplet ordering */
12 Sort the droplets in ascending order of their min pivot values.
of source and destination. The result of our proposed method for above
13 while droplet list is not null do
example is presented in Fig. 6.
14 if Multiple droplets having same min Pivot value then
4.4. Route compaction 15 Select one or multiple pivot cells based on proposed
greedy based heuristic.
16 else
In Final phase, compaction is performed to generate a parallel
17 Select the minimum valued pivot cell.
moving sequence. Here, every droplet is moved a single cell at a time
along its route. If two droplets violets fluidic constraints, deadlock 18 end if
occurs. To break the deadlock, stall and detour is used. When all the 19 end while
droplets reached the final destination, the parallel movement sequence 20 /* Final route generation */
is generated. 21 Generate two ants from each selected pivot cell i.e. one
towards source cell other towards target cell.
4.5. Algorithm 22 Store the sub-route in the header of the first ant to reach the
source or destination cell.
Let us consider a 𝑝 × 𝑞 biochip consisting of set of blockages and free 23 Append both sub-routes to generate the final route.
cells. A set of droplet D = {FD1 , FD2 , . . . , FD𝑛 } where each droplet FD𝑖
is a source and destination pair. Our proposed algorithm is presented
in Algorithm 1. In first phase, the whole chip is explored. The route 5. Experimentation and result analysis
exploration is presented in step 2–5. The 𝑓 𝑜𝑟() loop in step 3, initiates
route exploration by generating two waveforms. It terminates when 5.1. System setup
no free cells are available for marking. The 𝑓 𝑜𝑟() loop in step 6 to
10, finds out all the pivot cells of every droplet. The pivot cells are We perform our experimentations in a 64-bit Windows 10 desktop
assigned values by summing up the iteration numbers. The pivot values PC, with 16 GB of RAM and an Intel Core i5™ CPU operating at 2.8 GHz.
are sorted in ascending order and the min pivot value cell is identified. We use UCR microfluidic static simulator [24] for our test runs. UCR
This generates a local ordering of pivot cells. In step 12, we sort the microfluidic static simulator is an open source simulator developed
pivot values of all droplets to generate the final droplet order. Step 13– in University of California, Riverside. It uses Java for its graphical
19, performs the route selection. First we check for multiple pivot cells interfaces and C++ for its back end operations. It has many routing
having min values. If so, we select one or multiple pivot cells based protocols built in to it. This saves implementation time for protocols.
on one of the proposed heuristics i.e. random, single greedy, multiple For scheduling phase, path scheduler [5] is used. For placement phase,
greedy and all greedy. Then, two ants are generated for every selected Krammer placer [8] is used for placement of various modules.

8
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Table 4
Number of shared cells in virtual topology
Benchmark Pan2011[9] Grissom2018 [24] Juarez2018[26] Swain2019 [12] Huang2020 [16] Total random Single greedy Multiple greedy All greedy
PCR 572 362 436 413 502 631 361 424 473
In-Vitro 1 614 487 547 486 523 581 504 592 551
In-Vitro 2 839 761 802 763 760 944 783 771 821
In-Vitro 3 1235 1368 1363 1304 1376 1637 1364 1470 1307
In-Vitro 4 2064 2021 2174 Fail 194 1941 1872 1974 1880
In-Vitro 5 3721 4362 4405 Fail 4027 3962 3405 3685 2563
Protein Split 1 946 1073 1102 1058 1136 1350 939 1028 1132
Protein Split 2 1927 2475 2351 1862 2057 3074 2374 2109 1890
Protein Split 3 5614 Fail 5573 Fail 6381 7498 6430 5376 6037
Protein Split 4 Fail Fail Fail Fail Fail 30876 19467 18573 15462
Protein Split 5 Fail Fail Fail Fail Fail 50211 48963 43620 40736
Protein Split 6 Fail Fail Fail Fail Fail 103732 124650 123368 102653
Protein Split 7 Fail Fail Fail Fail Fail Fail 340728 356003 308679

5.2. Benchmarks total random methods gave best result. For protein split 1 benchmark,
Grissom router [24], Swain router [12] and Huang Router [16] have
We have used 13 popular publicly available benchmarks [1,7]. near equal run time. Similarly Grissom Router [24] and single greedy
These benchmarks are grouped into three families such as PCR, In-vitro method have near optimal value for Protein split 2 benchmark. The
diagnostics and a protein assay. For mixing operations, two 2 × 4 mixer LAT value for large assays is given Fig. 7(b). Juárez router [26] and
modules with time period of 3 s is used. Similarly for dilution, we use single greedy have same LAT value for Protein split 3. For Protein
the same 2 × 4 mixture modules with time cycle of 5 s. We also have split 4, Grissom router [24] has failed to run it, Pan router [9] have
one 2 × 4 detector and one 2 × 4 heater. They are used for detection and highest LAT value and Single greedy method have lowest LAT value.
heating operations in - vitro and protein assays. The droplet generation Pan router [9], Grissom Router [24], Juárez router [26] fails to run
rate or actuation frequency is set to 100 Hz [7]. Protein split 5,6 and 7 benchmarks. Multiple greedy method [12] gives
best LAT value for protein split 5. Huang router [16] fails to run Protein
5.3. Parameter split 6 and 7 benchmarks. All greedy methods gives optimal LAT value
for Protein split 6 and 7.
We have considered two architectures i.e. free and virtual topology Next we execute our proposed method in virtual topology, with 10%
for performance analysis. Free topology simulates ideal scenario with blockages. The result of our simulation is presented in Table 2. The
no blockages. Virtual topology emulates realistic scenarios with block- graphical presentation is given in Fig. 8. For PCR, Grissom Router [24],
ages. The dimension of biochip is 19 × 21. For comparative study with Huang Router [16] has best LAT value. Both single and multiple greedy
other popular routing protocols two parameters such as latest arrival method have approximately lowest LAT for In-vitro 1, 2 and 3. Juárez
time (LAT) and number of shared cells are chosen. router [26] have best latest arrival time for In-vitro 4. Swain router [12]
failed to run In-vitro 4 and 5 as it is tailored for free topology, and
5.4. Result and discussion may have spent more time in exploring route frequently. Grissom
router [24] gives smallest LAT value for both protein split 1 and 2,
We have performed comparison with five popular protocols. Three but fails to run protein split 3–7 benchmarks. Swain router [12] also
traditional methods such as Grissom et al. [24], Swain et al. [12] can’t execute protein split 3–7 benchmarks. Multiple greedy method
and Huang et al. [16] and two swarm intelligence based methods i.e gives smallest LAT value for protein split 4 & 5 benchmarks. All
Pan et al. [9] (ant colony optimization) and Juárez et al. [26] (gentic greedy method have lowest latest arrival time for protein split 6 & 7
algorithm). Grissom et al. [24] is defined for virtual topology while benchmarks.
Swain et al. is tailor-made for free topology. For virtual topology, The following observations are made from the above result. The
we randomly choose 10% of cells as blockages. The distribution of execution time and latest arrival time increases as the density of bench-
blockages are made such that the graph don’t become disconnected. mark increases. For smaller assays, our proposed method performs
The benchmarks are grouped into two groups based on the number averagely i.e. near to optimal value. For large group, our proposed
of nodes in their DAG i.e. small and large group. Small group has the method outshine all other methods. Total random heuristics gives
assays that have ≤100 nodes, while large group have benchmarks that different result for every test run. In order to stabilize, we executed
have more than 100 nodes. All benchmarks have been executed 10 every benchmark 10 times and calculate average to get the final result.
times. We have taken the average of result obtained in these test runs For free topology, total random heuristics gives best result for most of
and presented in tabular form. the benchmarks in free topology. The possible reason is the absence
of the overhead of selecting one or multiple pivot cells. There is an
5.4.1. Latest arrival time improvement of 7.32% and 18.05% in latest arrival time in free and
The first goal of our analysis is to study the execution time of virtual topology respectively.
our proposed method. For this, latest arrival time (LAT) metric is
selected. LAT refers to the time period at which the last droplet of an 5.4.2. Number of shared cells
assay reaches its destination. Here, we measure LAT in milliseconds. Our third goal is to identify the cells that are most frequently used
We first executed our benchmarks in free topology and the result is in the biochip. Identifying these cells help to improve the reliability
presented in Table 1. The LAT for small benchmarks are represented of the biochip. These heavily used cells tend to deteriorate and may
in Fig. 7(a). The smallest benchmark PCR, is successfully ran by all behave maliciously or become a blockage after a specified number
the methods. Our proposed method using total random heuristics gives of usage. These heavily cells can be replaced in due time to increase
the best LAT. Grissom router [24] have best LAT for PCR and In- in the reliability of biochip. Hence, we define a new heuristic called
vitro 1 benchmarks as it uses fast Roy router [10]. Swain router [12] 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠ℎ𝑎𝑟𝑒𝑑 𝑐𝑒𝑙𝑙𝑠 for this purpose. Number of shared cells specifies
performs best for In-vitro 2 and 3. For In-vitro 4 multiple greedy and the number of cells that have been common in routes of the multiple

9
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Fig. 9. Number of shared cells in free topology (a) Benchmarks PCR,In-Vitro 1–5 and Protein split 1–2 (b) Protein split 3–7.

number of droplet. Final route map contains both disjoint and shared router [26] fails to run protein split 5, 6 & 7. Swain router [12] able
routes. If the number of disjoint paths is more, the number of shared to run all the benchmarks. It has best result for PCR, In-vitro 1, protein
cells will be less and vice versa. The heuristic 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠ℎ𝑎𝑟𝑒𝑑 𝑐𝑒𝑙𝑙𝑠 split 1 & 2. Huang router [16] have least number of shared cells for
can specify the state of congestion in the biochip. We have executed In-vitro 2 & protein split 1, but fails to run protein split 6 & 7. Next
all the benchmarks in both free and virtual topologies. In order to we execute all the benchmarks in virtual topology. The simulation is
find the number of shared cells, we use a counter value to store
given in Table 4. The graphical representation is shown in Fig. 10.
the number of usage of cell. We incorporated this counter to all the
Pan router [9] don’t have optimal number of shared cells for any
routers considered for performance analysis. First, we execute all the
benchmarks. Grissom router [24] have minimum number of shared
benchmarks in free topology. The result of our simulation is listed in
Table 3. The graphical representation is given in Fig. 9. Pan Router [9] cells for PCR and In-vitro 1. Swain router [12] have least number of
able to run all assays except protein split 5, 6 & 7. It has minimum shared cells for In-vitro 1 & 2, protein split 2. Huang router [16] have
number of shared cells In-vitro 2 benchmark. Grissom router [24] & lowest value for In-vitro 2. Pan router [9], Grissom router [24], Juárez
Juárez router [26] have optimal result for protein split 1 & 2 and Router [26], Swain router [12] & Huang router [16] fails to run protein
In-vitro 5 respectively. Similarly, both Grissom router [24] & Juárez split 4, 5, 6, & 7 benchmarks. Our proposed methods able to run all the

10
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Fig. 10. Number of shared cells in virtual topology (a) Benchmarks PCR,In-Vitro 1–5 and Protein split 1–3 (b) Protein split 4–7.

benchmarks. Single greedy method have optimal result for PCR & In- compaction, detouring is performed faster in our proposed method than
vitro 3 benchmarks. Multiple greedy method have minimal number of other method. We store multiple sub-routes in both source and destina-
shared cell for In-vitro 2. All greedy method gives the best result for tion cells. Finding alternative routes for detouring is much simpler that
In-vitro 5, protein split 2, 4, 5, 6 & 7. re-exploration of routes used in other methods. All the greedy methods
The following observations are derived from above simulation re- consider every the routes and select the best route. This leads best result
sult. The number of shared cells increases as size of the benchmark for denser benchmarks. Mathematically, we obtain an improvement of
increases. This is expected as the number of droplets increases, the 8.43% and 19.70% in number of shared cells for all benchmarks.
number of free cell reduces. This leads to less disjoint paths and
increase in number of shared cells. Swain router [12] is designed for 6. Analysis of proposed algorithm
free topology so failed for most benchmarks and performs poorly on
average. Our proposed method able to run all benchmarks because, To explore an 𝑎 × 𝑏 biochip, route exploration phase will take 𝑂(𝑎 ∗
we explore the whole chip in first phase. In case of failure in route 𝑏) time for a single droplet. Since, the total number of available cells is a

11
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

fixed, So it will take constant time and the time required is 𝑂(1). So for [7] K. O’neal, D. Grissom, P. Brisk, Resource-constrained scheduling for digital
n droplets, it will take 𝑂(𝑛). In next phase, the biochip was partitioned microfluidic biochips, J. Emerg. Technol. Comput. Syst. 14 (1) (2017) http:
into a bipartite graph, and pivot cells are identified and then they are //dx.doi.org/10.1145/3093930.
[8] D.T. Grissom, P. Brisk, Fast online synthesis of digital microfluidic biochips,
sorted. The pivot cell identification process, in worst case will explore
IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst. 33 (3) (2014) 356–369,
all the cells, which takes O(1). We are using randomized quick sort to http://dx.doi.org/10.1109/TCAD.2013.2290582.
sort the pivot values which have a runtime of 𝑂(𝑛𝑙𝑜𝑔𝑛) and selection [9] I. Pan, P. Dasgupta, H. Rahaman, T. Samanta, Ant colony optimization based
of minimum pivot cell takes O(1). So the pivot cell identification and droplet routing technique in digital microfluidic biochip, in: 2011 International
sorting and min pivot cell will take 𝑂(𝑛𝑙𝑜𝑔𝑛) time. The selection phase Symposium on Electronic System Design, 2011, pp. 223–229, http://dx.doi.org/
takes 𝑂(1) for randomized method for every droplet, so total time taken 10.1109/ISED.2011.23.
[10] P. Roy, H. Rahaman, P. Dasgupta, A novel droplet routing algorithm for
will be 𝑂(𝑛). For all other selection methods will also take 𝑂(𝑛). The
digital microfluidic biochips, in: Proceedings of the 20th Symposium on Great
time taken by the ants to reach their destination is 𝑂(𝑛) for every Lakes Symposium on VLSI, GLSVLSI ’10, Association for Computing Machinery,
droplet, so in total it will take 𝑂(𝑛2 ). Collecting the sub-routes and New York, NY, USA, 2010, pp. 441–446, http://dx.doi.org/10.1145/1785481.
appending to generate the final route will take 𝑂(1) time. So the total 1785583.
time taken by our proposed algorithm is 𝑂(𝑛2 ). [11] K. Bohringer, Towards optimal strategies for moving droplets in digital microflu-
idic systems, in: IEEE International Conference on Robotics and Automation,
2004. Proceedings, Vol. 2, ICRA ’04. 2004, 2004, pp. 1468–1474 Vol.2, http:
7. Conclusion and future work
//dx.doi.org/10.1109/ROBOT.2004.1308031.
[12] K.R. Jyotiranjan Swain, S. Pyne, A space efficient greedy droplet routing for
Routing in digital microfluidic biochip is an challenging task. The digital microfluidics biochip, in: 2019 23rd International Symposium, VLSI
goal is to move all the droplets from their source to destination cell. Design and Test, Vol. 235, 2019, pp. 131–139, http://dx.doi.org/10.1007/978-
We need to move maintain fluidic constraints for every step of the 981-32-9767-8_9.
[13] J. Swain, R. Kolluri, S. Pyne, A proactive wash droplet routing for digital
movement. The droplet routing problem belongs to NP hard family.
microfluidics biochip, in: 2020 24th International Symposium on VLSI Design
In this article, we proposed a greedy bidirectional droplet routing, and Test, VDAT, 2020, pp. 1–4, http://dx.doi.org/10.1109/VDAT50263.2020.
that gives an approximate optimal solution. Initially the whole chip is 9190390.
explored. Then, pivot cells are identified and pivot values are assigned. [14] T. Xu, K. Chakrabarty, A cross-referencing-based droplet manipulation method
The pivot cells are sorted in an ascending order. The pivot cells are for high-throughput and pin-constrained digital microfluidic arrays, in: Pro-
then selected using one of the four greedy route selection heuristics. For ceedings of Design Automation and Test in Europe, DATE, April 2007, pp.
552–557.
every droplets routes are selected and the final route map is generate
[15] P.-H. Yuh, S. Sapatnekar, C.-L. Yang, Y.-W. Chang, A progressive-ILP based
using route compaction. To handle deadlock stalls and detouring is routing algorithm for cross-referencing biochips, in: 2008 45th ACM/IEEE Design
used to generate the parallel moving sequence. Simulation result shows Automation Conference, 2008, pp. 284–289.
improvement in both parameters i.e. latest arrival time and number of [16] Z. Huang, X. Bai, T. Lan, X. Li, G. Lin, Unified contamination-aware routing
shared cells. In future, we plan to study the congestion problem using method considering realistic washing capacity constraint in digital microfluidic
various swarm intelligence algorithm to solve droplet routing problem biochips, IEEE Access 8 (2020) 192867–192879.
[17] Y. Zhao, K. Chakrabarty, Cross-contamination avoidance for droplet routing, in:
in digital microfluidic biochip.
Design and Testing of Digital Microfluidic Biochips, Springer New York, New
York, NY, 2013, pp. 27–55, http://dx.doi.org/10.1007/978-1-4614-0370-8_2.
Declaration of competing interest [18] K. Singha, T. Samanta, H. Rahaman, P. Dasguptay, Method of droplet routing
in digital microfluidic biochip, in: Proceedings of 2010 IEEE/ASME International
The authors declare that they have no known competing finan- Conference on Mechatronic and Embedded Systems and Applications, 2010, pp.
251–256, http://dx.doi.org/10.1109/MESA.2010.5552059.
cial interests or personal relationships that could have appeared to
[19] P. Roy, H. Rahaman, R. Bhattacharya, P. Dasgupta, A best path selection based
influence the work reported in this paper. parallel router for DMFBs, in: 2011 International Symposium on Electronic
System Design, 2011, pp. 176–181, http://dx.doi.org/10.1109/ISED.2011.33.
Data availability [20] P. Roy, H. Rahaman, P. Dasgupta, Two-level clustering-based techniques for
intelligent droplet routing in digital microfluidic biochips, Integration 45 (3)
(2012) 316–330, http://dx.doi.org/10.1016/j.vlsi.2011.11.006, Special Issue of
No data was used for the research described in the article.
GLSVLSI 2011: Current Trends on VLSI and Ultra Low-Power Design, URL
https://www.sciencedirect.com/science/article/pii/S0167926011000952.
References [21] P. Roy, P. Howladar, R. Bhattacharjee, H. Rahaman, P. Dasgupta, A new
cross contamination aware routing method with intelligent path exploration in
[1] F. Su, W. Hwang, K. Chakrabarty, Droplet routing in the synthesis of digital digital microfluidic biochips, in: 2013 8th International Conference on Design
microfluidic biochips, in: Proceedings of the Design Automation Test in Eu- Technology of Integrated Systems in Nanoscale Era, DTIS, 2013, pp. 50–55,
rope Conference, vol. 1, 2006, pp. 1–6, http://dx.doi.org/10.1109/DATE.2006. http://dx.doi.org/10.1109/DTIS.2013.6527777.
244177. [22] I. Pan, T. Samanta, Weighted optimization of various parameters for droplet
[2] E.J. Griffth, S. Akella, M.K. Goldberg, Performance characterization of a reconfig- routing in digital microfluidic biochips, in: S.M. Thampi, A. Abraham, S.K. Pal,
urable planar-array digital microfluidic system, in: K. Chakrabarty, J. Zeng (Eds.), J.M.C. Rodriguez (Eds.), Recent Advances in Intelligent Informatics, Springer
Design Automation Methods and Tools for Microfluidics-Based Biochips, Springer International Publishing, 2014, pp. 131–139.
Netherlands, Dordrecht, 2006, pp. 329–356, http://dx.doi.org/10.1007/1-4020- [23] R. Bhattacharya, P. Roy, H. Rahaman, Homogeneous droplet routing in DMFB: An
5123-9_13. enhanced technique for high performance bioassay implementation, Integration
[3] E. Griffth, S. Akella, Coordinating multiple droplets in planar array digital
60 (2018) 74–91, http://dx.doi.org/10.1016/j.vlsi.2017.08.001, URL https://
microfluidics systems, Int. J. Robot. Res. 24 (11) (2005) 933–949, http://dx.
www.sciencedirect.com/science/article/pii/S016792601730264X.
doi.org/10.1177/0278364905059067.
[24] C. Curtis, D. Grissom, P. Brisk, A compiler for cyber-physical digital microfluidic
[4] M. Cho, D.Z. Pan, A high-performance droplet routing algorithm for digital
biochips, in: CGO 2018, Association for Computing Machinery, New York, NY,
microfluidic biochips, IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst. 27
USA, 2018, pp. 365–377, http://dx.doi.org/10.1145/3168826.
(10) (2008) 1714–1724, http://dx.doi.org/10.1109/TCAD.2008.2003282.
[25] J. Swain, S. Pyne, Deadlock detection in digital microfluidics biochip droplet
[5] D. Grissom, P. Brisk, Path scheduling on digital microfluidic biochips, in:
routing, in: 2018 22rd International Symposium, VLSI Design and Test, Vol. 892,
Proceedings of the 49th Annual Design Automation Conference, DAC ’12,
Association for Computing Machinery, New York, NY, USA, 2012, pp. 26–35, 2018, pp. 242–253, http://dx.doi.org/10.1007/978-981-13-5950-7_21.
http://dx.doi.org/10.1145/2228360.2228367. [26] J. Juárez, C.A. Brizuela, I.M. Martínez-Pérez, An evolutionary multi-objective
[6] T.-W. Huang, T.-Y. Ho, A fast routability- and performance-driven droplet optimization algorithm for the routing of droplets in digital microfluidic biochips,
routing algorithm for digital microfluidic biochips, in: 2009 IEEE International Inform. Sci. 429 (2018) 130–146, http://dx.doi.org/10.1016/j.ins.2017.11.011.
Conference on Computer Design, 2009, pp. 445–450, http://dx.doi.org/10.1109/
ICCD.2009.5413119.

12
J. Swain and S. Pyne Microprocessors and Microsystems 98 (2023) 104774

Jyotiranjan Swain is currently pursuing Doctor of Phi-


losophy from National Institute of Technology, Rourkela, Sumanta Pyne is an Assistant Professor at department
India. His research area is Digital microfluidic biochip. He of Computer Science and Engineering, National Institute
has Master of Technology (M.Tech2014) from University of Technology, Rourkela, India. His research interest in-
of Calicut, Kerala, India and completed his Bachelor of clude Low power VLSI Design, Digital microfluidic biochip,
Technology (B.Tech2007) from University College of Engi- Computer Networks. He has Ph.D. from Indian Institue of
neering (Currently known as Veer Surinder Sai University Technology, Kharagpur, India. He completed his Master of
of Technology), Sambalpur, India. Engineering (M.E.) from Bengal Engineering and Science
University, West Bengal, INDIA. (presently Indian Insti-
tute of Engineering Science and Technology) and B.Tech
from West Bengal University of Technology, Kolkata, IN-
DIA. (presently Maulana Abul Kalam Azad University of
Technology).

13

You might also like