Research Papers: Techniques and Procedures

Automated Design for Microfluid Flow Sculpting: Multiresolution Approaches, Efficient Encoding, and CUDA Implementation

[+] Author and Article Information
Daniel Stoecklein, Michael Davies

Department of Mechanical Engineering,
Iowa State University,
Ames, IA 50011

Nadab Wubshet

Department of Physics,
Augustana University,
Sioux Falls, SD 57197

Jonathan Le

Hoover High School,
Des Moines, IA 50310

Baskar Ganapathysubramanian

Associate Professor
Department of Mechanical Engineering,
Iowa State University,
Ames, IA 50011
e-mail: baskarg@iastate.edu

1Corresponding author.

Contributed by the Fluids Engineering Division of ASME for publication in the JOURNAL OF FLUIDS ENGINEERING. Manuscript received June 16, 2016; final manuscript received October 2, 2016; published online January 20, 2017. Assoc. Editor: Kwang-Yong Kim.

J. Fluids Eng 139(3), 031402 (Jan 20, 2017) (11 pages) Paper No: FE-16-1373; doi: 10.1115/1.4034953 History: Received June 16, 2016; Revised October 02, 2016

Sculpting inertial fluid flow using sequences of pillars is a powerful method for flow control in microfluidic devices. Since its recent debut, flow sculpting has been used in novel manufacturing approaches such as microfiber and microparticle design, flow cytometry, and biomedical applications. Most flow sculpting applications can be formulated as an inverse problem of finding a pillar sequence that results in a desired fluid transformation. Manual exploration and design of pillar sequences, while useful, have proven infeasible for finding complex flow transformations. In this work, we extend our automated optimization framework based on genetic algorithms (GAs) to rapidly design micropillar sequences that can generate arbitrary user-defined fluid flow transformations. We design the framework with the following properties: (a) a parameter encoding that respects locality to ensure fast convergence and (b) a multiresolution approach that accelerates convergence while maintaining accuracy. The framework also utilizes graphics processing unit (GPU) architecture via NVIDIA's CUDA for function evaluations. We package this framework in a user-friendly and freely available software suite that enables the larger microfluidics community to utilize these developments. We also demonstrate the framework's capability to rapidly design arbitrary fluid flow shapes across multiple microchannel aspect ratios.

Copyright © 2017 by ASME
Your Session has timed out. Please sign back in to continue.


Amini, H. , Sollier, E. , Masaeli, M. , Xie, Y. , Ganapathysubramanian, B. , Stone, H. A. , and Di Carlo, D. , 2013, “ Engineering Fluid Flow Using Sequenced Microstructures,” Nat. Commun., 4, pp. 1826–1834. [CrossRef] [PubMed]
Stoecklein, D. , Wu, C.-Y. , Owsley, K. , Xie, Y. , Di Carlo, D. , and Ganapathysubramanian, B. , 2014, “ Micropillar Sequence Designs for Fundamental Inertial Flow Transformations,” Lab Chip, 14(21), pp. 4197–4204. [CrossRef] [PubMed]
Lepchev, D. , and Weihs, D. , 2010, “ Low Reynolds Number Flow in Spiral Microchannels,” ASME J. Fluids Eng., 132(7), p. 071202. [CrossRef]
Zhou, T. , Xu, Y. , Liu, Z. , and Joo, S. W. , 2015, “ An Enhanced One-Layer Passive Microfluidic Mixer With an Optimized Lateral Structure With the Dean Effect,” ASME J. Fluids Eng., 137(9), p. 091102. [CrossRef]
Wang, Y. , Zhe, J. , Dutta, P. , and Chung, B. T. , 2007, “ A Microfluidic Mixer Utilizing Electrokinetic Relay Switching and Asymmetric Flow Geometries,” ASME J. Fluids Eng., 129(4), pp. 395–403. [CrossRef]
Stoecklein, D. , Wu, C.-Y. , Kim, D. , Di Carlo, D. , and Ganapathysubramanian, B. , 2016, “ Optimization of Micropillar Sequences for Fluid Flow Sculpting,” Phys. Fluids, 28(1), pp. 53–65. [CrossRef]
Nunes, J. K. , Wu, C. Y. , Amini, H. , Owsley, K. , Di Carlo, D. , and Stone, H. A. , 2014, “ Fabricating Shaped Microfibers With Inertial Microfluidics,” Adv. Mater., 26(22), pp. 3712–3717. [CrossRef] [PubMed]
Paulsen, K. S. , Di Carlo, D. , and Chung, A. J. , 2015, “ Optofluidic Fabrication for 3D-Shaped Particles,” Nat. Commun., 6, pp. 6976–6985. [CrossRef] [PubMed]
Wu, C.-Y. , Owsley, K. , and Di Carlo, D. , 2015, “ Rapid Software-Based Design and Optical Transient Liquid Molding of Microparticles,” Adv. Mater., 27(48), pp. 7970–7978. [CrossRef] [PubMed]
Paulsen, K. S. , and Chung, A. J. , 2016, “ Non-Spherical Particle Generation From 4D Optofluidic Fabrication,” Lab Chip, 16(16), pp. 2987–2995. [CrossRef] [PubMed]
Sollier, E. , Amini, H. , Go, D. , Sandoz, P. , Owsley, K. , and Di Carlo, D. , 2015, “ Inertial Microfluidic Programming of Microparticle-Laden Flows for Solution Transfer Around Cells and Particles,” Microfluid. Nanofluid., 19(1), pp. 53–65. [CrossRef]
Foli, K. , Okabe, T. , Olhofer, M. , Jin, Y. , and Sendhoff, B. , 2006, “ Optimization of Micro Heat Exchanger: CFD, Analytical Approach and Multi-Objective Evolutionary Algorithms,” Int. J. Heat Mass Transfer, 49, pp. 1090–1099. [CrossRef]
Lee, Y.-T. , Ahuja, V. , Hosangadi, A. , and Ebert, M. , 2010, “ Shape Optimization of a Multi-Element Foil Using an Evolutionary Algorithm,” ASME J. Fluids Eng., 132(5), p. 051401. [CrossRef]
Ahuja, V. , Hartfield, R. J. , and Shelton, A. , 2014, “ Optimization of Hypersonic Aircraft Using Genetic Algorithms,” Appl. Math. Comput., 242, pp. 423–434.
Lyutov, A. E. , Chirkov, D. V. , Skorospelov, V. A. , Turuk, P. A. , and Cherny, S. G. , 2015, “ Coupled Multipoint Shape Optimization of Runner and Draft Tube of Hydraulic Turbines,” ASME J. Fluids Eng., 137(11), p. 111302. [CrossRef]
Holland, J. H. , 1992, “ Genetic Algorithms,” Sci. Am., 267(1), pp. 66–71. [CrossRef]
Kim, D. S. , Jung, D. H. , and Kim, Y. Y. , 2008, “ Multiscale Multiresolution Genetic Algorithm With a Golden Sectioned Population Composition,” Int. J. Numer. Methods Eng., 74(3), pp. 349–367. [CrossRef]
Babbar, M. , and Minsker, B. S. , 2006, “ Groundwater Remediation Design Using Multiscale Genetic Algorithms,” J. Water Resour. Plann. Manage., 132(5), pp. 341–350. [CrossRef]
Saito, M. , and Matsumoto, M. , 2008, “ SIMD-Oriented Fast Mersenne Twister: A 128-bit Pseudorandom Number Generator,” Monte Carlo and Quasi-Monte Carlo Methods 2006, A. Keller , S. Heinrich , and H. Niederreiter , eds., Springer, Berlin, pp. 607–622.


Grahic Jump Location
Fig. 1

Illustration of a three-pillar flow sculpting device with a three-channel inlet. The middle channel contains the fluid that is “sculpted” by the micropillar sequence. In this sequence, each pillar has the same normalized diameter D/w and is at the same location y/w in a microchannel of height h. With sufficient interpillar spacing (≥6D/w, for Re ≤ 40), the fluid deformation from each pillar has time to saturate before reaching the next pillar. The cross-sectional images showing the inlet flow condition and fluid deformations after each pillar are shown below the microchannel. These images are created by the same simulation method as in Ref. [6], which is used in this work as well.

Grahic Jump Location
Fig. 2

A comparison of the workflow necessary for pillar programming design based on the baseline GA framework (a) and the one presented in this paper (b). Previously, the user needed to specify both a desired flow shape (a-i) and the inlet flow configuration (a-ii). This could require multiple user-guided iterations on inlet configurations before a solution is found (a-iii). Now that the inlet is a part of the GA chromosome the user simply supplies the GA framework their desired image (b-i) and the number of possible channels at the microchannel inlet (b-ii). The solution chromosome (b-iii) contains the optimized inlet design as well as the pillar sequence.

Grahic Jump Location
Fig. 3

(a) Schematic for the inlet design portion of the chromosome, where the number of values in the chromosome represents the number of channels being joined at the inlet. The values are binary, where 0 represents nontracked fluid and 1 is the fluid of interest being sculpted. (b) The ℝ-chromosome design uses two values for each pillar: diameter and location. ((c) and (d)) These values are given equal probability by their bounds in the GA, and interpreted in real-time to select the nearest available precomputed transition matrix. The entire chromosome for the design as shown would be [0, 1, 0, 1, 0, 0.375, 0.0, 0.375, 0.25, 0.375, −0.375].

Grahic Jump Location
Fig. 4

(a) Graphical user interface (GUI) for using the GA framework. Note the drawing canvas (top of GUI) with enforced channel symmetry, GA parameters (lower left), and inlet/sequence design for a microfluidic device (bottom). (b) UML class diagram of custom GA framework.

Grahic Jump Location
Fig. 5

Examples of randomly generated target images and results from the GA framework (without using FFT postprocessing in the fitness function), with their fitness values. Sculpted fluid is shown in black. The target images were all made with the same inlet design of [0, 0, 1, 0, 0] (see Fig. 3(a)).

Grahic Jump Location
Fig. 6

Results for case 1, which tested different transition matrix resolutions for accuracy and speed. Sixty different flow shape images were generated using transition matrices from sequences determined by the quasi-random Sobol set, and a fixed inlet configuration. Each target image was blurred and thresholded to mimic a realistic user-supplied design. Framework accuracy (a) and runtime (b) are shown for each transition matrix size, for all 60 target images. Note that the fitness function accuracy (a) has been normalized relative to the baseline (NY = 801).

Grahic Jump Location
Fig. 7

Results for case 2. (a) Use of a multiresolution fitness function (NY(ML)=51, NY(MH)=601) resulted in similar optimal fitness to the baseline of NY = 601. (b) Use of NY = 51 transition matrices for the first 50 generations resulted in substantially decreased runtime.

Grahic Jump Location
Fig. 8

Results for case 3, which tests the ℝ-chromosome against the integer-proxy ℤ-chromosome. The inlet design is fixed in the chromosome, thus isolating the test to the choice of pillar sequence encoding. (a) The ℝ-chromosome results in comparable accuracy in optimal fitness for each target image. Here, the fitness for the best result from each GA framework search is reported for all 60 images. (b) Use of the ℝ-chromosome shows faster convergence in the GA with a mean of 63 generations per GA. Note that the ℤ-chromosome, with a mean of 111 generations, had a significant proportion of GA searches terminate at the generation limit of 200. (c) Fewer generations translate into less framework runtime, with a median of ≈7000 s for the ℤ-chromosome, and ≈3600 s for the ℝ-chromosome.

Grahic Jump Location
Fig. 9

Results for case 4, which tested the GA's ability to design the inlet flow configuration. (a) Optimal fitness results for 60 target flow shapes with randomly generated pillar sequences and inlet flow configurations. Sculpted fluid is shown in black. The best match by fitness, with f = 5.43, is shown in (b), while the worst match, with f = 45.56, is shown in (c). Although the discrepancies between the target fluid shape and the GA result in (c) are clearly discernible, the overall design has still been found.

Grahic Jump Location
Fig. 10

Result for case 5, which compares performance of the custom GA to the matlab implementation. (a) Comparison of optimal fitness values for each of the 60 target images, which are identical to those used in case 1. Unlike case 1, the portion of the chromosome which governs the inlet design was accessible to the GA, thus making the problem more difficult. Note that the custom GA accuracy matches, and in some cases exceeds that of the matlab GA. (b) Runtime distributions for the custom GA and matlab GA. The high complexity of the problem compared to case 1 (designing the inlet as well as the pillar sequence) requires additional generations for convergence, which increases the runtime.

Grahic Jump Location
Fig. 11

Results for case 6, which uses the new GA framework for hand-drawn flow shapes of the letters that spell IOWA, for microchannel aspect ratio h/w = 0.25

Grahic Jump Location
Fig. 12

(a) A simple illustration of how reverse advection is used to create an advection map (b) that contains uniformly distributed information for fluid displacement at the microchannel outlet. The advection map can then be converted into a column-stochastic transition matrix (c), for which every row and column represents a fluid element in the two-dimensional (2D) cross section shown in (a) for the inlet and outlet, respectively. Thus, the displacement for a fluid element that would otherwise be calculated by streamtracing through a 3D domain (d) can be computed using only matrix multiplication (e). The inlet fluid states are discretized in the same way as the transition matrix (e-i), and then reshaped to form a row vector (e-ii). This is then multiplied by a transition matrix (e-iii), which produces the outlet fluid states as a row vector (e-iv). This can then be reshaped into the 2D representation of the domain (e-v).

Grahic Jump Location
Fig. 13

Additional results for case 6, which uses the new GA framework for hand-drawn flow shapes of the letters that spell IOWA, for microchannel aspect ratios h/w = 0.5 (a) and h/w = 1.0 (b)




Some tools below are only available to our subscribers or users with an online account.

Related Content

Customize your page view by dragging and repositioning the boxes below.

Related Journal Articles
Related eBook Content
Topic Collections

Sorry! You do not have access to this content. For assistance or to subscribe, please contact us:

  • TELEPHONE: 1-800-843-2763 (Toll-free in the USA)
  • EMAIL: asmedigitalcollection@asme.org
Sign In