pith. sign in

arxiv: 2606.25569 · v1 · pith:SVIWVDUYnew · submitted 2026-06-24 · ⚛️ physics.flu-dyn · cs.CE

VesNet: Neural network accelerated solver for simulating Stokesian vesicle suspensions

Shan Zhong , Gokberk Kabacaoglu , George Biros This is my paper

Pith reviewed 2026-06-25 20:50 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn cs.CE
keywords vesicle suspensionsStokes flowneural networkboundary integral methodGPU accelerationfluid-structure interactiondeformable particleslubrication forces
0
0 comments X

The pith

A neural network can approximate vesicle self-interactions to deliver over 100x faster simulations of Stokesian vesicle suspensions while matching the original dynamics.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents VesNet, a hybrid solver for two-dimensional vesicle suspensions in Stokes flow. It replaces the expensive computation of each vesicle's self-interaction, background flow coupling, and short-range lubrication forces with a neural network approximation. Standard boundary integral techniques continue to handle boundary reparameterization and far-field interactions between vesicles. The resulting GPU implementation runs more than 100 times faster than a multithreaded CPU version of the full solver and about 5 times faster than a GPU version of the original code. The accelerated runs still reproduce single-vesicle phase behavior, pair interactions, and collective motion in Taylor-Green and Poiseuille flows.

Core claim

VesNet accelerates two-dimensional vesicle suspension simulations by approximating vesicle self interactions, including background flow coupling and short-range lubrication forces, while retaining conventional modules for boundary reparameterization and far-field hydrodynamics. A GPU-accelerated implementation achieves over 100x speedup compared to a multithreaded MATLAB CPU boundary integral solver and about 5x relative to its GPU counterpart. VesNet accurately captures key dynamics, including single-vesicle phase behavior, pair interactions, and large-scale suspensions in Taylor-Green and Poiseuille flows, enabling efficient simulations of thousands of vesicles on modest computational reso

What carries the argument

VesNet, the hybrid framework that substitutes a neural network for the self-interaction terms of each vesicle while keeping conventional modules for boundary reparameterization and far-field hydrodynamics.

If this is right

  • Thousands of vesicles can be simulated on standard hardware instead of requiring large clusters.
  • Collective behaviors in prescribed flows such as Taylor-Green and Poiseuille become practical to explore at scale.
  • Pairwise interactions and single-vesicle phase transitions remain faithful to the underlying boundary integral formulation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same substitution of neural approximations for local interactions could be tested on other deformable particles in Stokes flow after retraining.
  • Extension to three dimensions would require generating and validating new training data from three-dimensional boundary integral runs.
  • Parameter studies over vesicle properties or flow strengths could now be performed at higher resolution or with more realizations.
  • pacs':['47.57.ef','47.63.Gd','47.11.-j'],

Load-bearing premise

The neural network approximation of vesicle self interactions, including background flow coupling and short-range lubrication forces, remains sufficiently accurate to reproduce the dynamics of the original boundary integral method without introducing errors that alter the reported behaviors.

What would settle it

A side-by-side run of identical initial conditions in VesNet and the full boundary integral solver that produces visibly different single-vesicle phase diagrams or pair-interaction trajectories would falsify the accuracy claim.

Figures

Figures reproduced from arXiv: 2606.25569 by George Biros, Gokberk Kabacaoglu, Shan Zhong.

Figure 1
Figure 1. Figure 1: 2,000 vesicles in Poiseuille flow (see sections [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Vesicle suspension in a free-space Taylor-Green flow. The left figure shows the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Near-singular integration schemes with the state-of-the-art numerical method (in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The architecture of a single block in IVNet for updating 1D hidden [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Main components of VesNet. (a) Given the vesicle configurations at the current [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Snapshots of an arbitrary shaped vesicle in a stationary fluid. Vesicle relaxes to [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Phase diagram of equilibrium vesicle shapes in Poiseuille flow. BIEM (on the [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Equilibrium lateral position in Poiseuille flow. Each figure corresponds to [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Two vesicles in shear flow. Each row shows snapshots from simulations. The [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Separation between two vesicles in shear flow. In the simulations, the [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Dense suspension in Taylor-Green flow with velocity ( [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Evaluation of the proposed near-singular integration scheme with five layers at [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The evolution of relative error in X of the relaxation network during 100 relaxation steps. The distribution of errors for 3200 vesicle samples is shown. Solid blue line shows the median error, while shaded regions indicate percentile spreads [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: A large relaxation error example (upper row) and a normal error example [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Error distribution of the self tension network. [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Error distributions of the advection network. The left figure shows the network [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Error distributions of the advective tensions. The left figure shows the network [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Error distributions of the near field networks, at different layers. Each figure [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
read the original abstract

Numerical simulation of deformable particle suspensions in Stokes flow is computationally expensive due to nonlinear fluid-structure interactions, evolving interfaces, and multiscale hydrodynamics. We present VesNet, a hybrid framework that accelerates two-dimensional vesicle suspension simulations by approximating vesicle self interactions, including background flow coupling and short-range lubrication forces, while retaining conventional modules for boundary reparameterization and far-field hydrodynamics. A GPU-accelerated implementation achieves over 100x speedup compared to a multithreaded MATLAB CPU boundary integral solver and about 5x relative to its GPU counterpart. VesNet accurately captures key dynamics, including single-vesicle phase behavior, pair interactions, and large-scale suspensions in Taylor-Green and Poiseuille flows, enabling efficient simulations of thousands of vesicles on modest computational resources.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript presents VesNet, a hybrid solver for 2D Stokesian vesicle suspensions that uses a neural network to approximate vesicle self-interactions (background flow coupling and short-range lubrication) while retaining conventional boundary-integral modules for far-field hydrodynamics and interface reparameterization. A GPU implementation is reported to deliver >100x speedup versus a multithreaded MATLAB CPU BIE solver and ~5x versus its own GPU counterpart, with claims that the method accurately reproduces single-vesicle phase behavior, pair interactions, and large-scale dynamics in Taylor-Green and Poiseuille flows.

Significance. If the NN surrogate for self-interactions can be shown to control long-time error accumulation and preserve qualitative dynamics, the hybrid approach would enable previously intractable simulations of O(10^3) vesicles on modest hardware, directly addressing the computational bottleneck in vesicle suspension studies. The decision to keep far-field and reparameterization steps exact is a sound architectural choice that limits the scope of approximation error.

major comments (2)
  1. [Abstract] Abstract: the headline accuracy claim (that VesNet 'accurately captures key dynamics' for single-vesicle, pair, and large-scale cases) is unsupported by any quantitative error metrics, area/volume conservation drift, or direct long-time trajectory comparisons against the reference BIE solver over O(10^3–10^4) steps; without these, it is impossible to verify that local surrogate discrepancies do not alter reported phase behavior or suspension statistics.
  2. [Abstract] The central performance claim rests on the NN approximation of self-interactions (including lubrication) remaining sufficiently accurate that the hybrid evolution reproduces the original BIE dynamics; the manuscript provides no a-posteriori error bounds, conservation properties, or accumulation analysis for this surrogate, leaving open the possibility that repeated pair interactions amplify small discrepancies.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point-by-point below and will revise the manuscript to strengthen the presentation of accuracy and error analysis.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline accuracy claim (that VesNet 'accurately captures key dynamics' for single-vesicle, pair, and large-scale cases) is unsupported by any quantitative error metrics, area/volume conservation drift, or direct long-time trajectory comparisons against the reference BIE solver over O(10^3–10^4) steps; without these, it is impossible to verify that local surrogate discrepancies do not alter reported phase behavior or suspension statistics.

    Authors: We agree that the abstract would benefit from quantitative support. The results section contains error metrics, conservation data, and trajectory comparisons; we will revise the abstract to report key quantitative measures (e.g., shape errors, area conservation drift, and long-time agreement) drawn from those sections. revision: yes

  2. Referee: [Abstract] The central performance claim rests on the NN approximation of self-interactions (including lubrication) remaining sufficiently accurate that the hybrid evolution reproduces the original BIE dynamics; the manuscript provides no a-posteriori error bounds, conservation properties, or accumulation analysis for this surrogate, leaving open the possibility that repeated pair interactions amplify small discrepancies.

    Authors: We will add an explicit subsection summarizing a-posteriori error bounds, conservation properties, and accumulation analysis for the NN surrogate over long trajectories to directly address concerns about error propagation in repeated interactions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical NN surrogate validated against reference BIE

full rationale

The VesNet framework trains a neural network on data generated by the conventional boundary-integral solver to approximate local self-interactions and lubrication, then couples it to unchanged far-field and reparameterization modules. Speedup and accuracy statements are direct empirical measurements on held-out single-vesicle, pair, and suspension test cases; none of the reported quantities is obtained by fitting a parameter to the same data and relabeling it a prediction. No uniqueness theorem, self-citation chain, or ansatz is invoked to force the architecture or the reported behaviors. The method remains externally falsifiable by comparison to the original BIE code.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so free parameters, axioms, and invented entities cannot be enumerated. The central claim rests on the unverified accuracy of the neural-network approximation of self-interactions.

pith-pipeline@v0.9.1-grok · 5659 in / 1140 out tokens · 19632 ms · 2026-06-25T20:50:12.421600+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 3 canonical work pages

  1. [1]

    and Yeung, D

    Zhang, Y. and Yeung, D. , title =. In Proc. ACM SIGKDD Conf. , year =

  2. [2]

    Newman, M. E. J. , title =

  3. [3]

    and Fortunato, S

    Lancichinetti, A. and Fortunato, S. , title =. Phys. Rev. E. , volume =. 2009 , pages =

    2009

  4. [4]

    Procedia IUTAM , volume =

    High-order adaptive time stepping for vesicle suspensions with viscosity contrast , author =. Procedia IUTAM , volume =

  5. [5]

    Li, Zongyi and Kovachki, Nikola and Azizzadenesheli, Kamyar and Liu, Burigede and Bhattacharya, Kaushik and Stuart, Andrew and Anandkumar, Anima , year =. Fourier. 2010.08895 , primaryclass =

    Pith/arXiv arXiv 2010

  6. [6]

    Learning Nonlinear Operators via

    Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zhongqiang and Karniadakis, George Em , year =. Learning Nonlinear Operators via. Nat Mach Intell , volume =

  7. [7]

    2024 , month = jan, journal =

    Derivative-. 2024 , month = jan, journal =

    2024

  8. [8]

    Choose a

    Cao, Shuhao , year =. Choose a. 2105.14995 , primaryclass =

    arXiv

  9. [9]

    2024 , month = jan, number =

    Kopani. 2024 , month = jan, number =. 2401.02016 , primaryclass =

    arXiv 2024

  10. [10]

    PNAS , volume =

    A vesicle bioreactor as a step toward an artificial cell assembly , author =. PNAS , volume =

  11. [11]

    Biomimetic nanoscale reactors and networks , author =. Ann. Rev. Phys. Chem. , volume =

  12. [12]

    Science , volume =

    Drug delivery systems: Entering the mainstream , author =. Science , volume =

  13. [13]

    Trends Biotechnol

    Engineering liposomes for drug delivery , author =. Trends Biotechnol. , volume =

  14. [14]

    Journal of Computational Physics , volume =

    A fast algorithm for particle simulations , author =. Journal of Computational Physics , volume =

  15. [15]

    Journal of Computational Physics , volume =

    Non-intrusive reduced order modeling of unsteady flows using artificial neural networks with application to a combustion problem , author =. Journal of Computational Physics , volume =

  16. [16]

    The viscosity of the blood in narrow capillary tubes , author =. Am. J. Physiol. , volume =

  17. [17]

    AIAA Journal , volume =

    Interpolation Method for Adapting Reduced-Order Models and Application to Aeroelasticity , author =. AIAA Journal , volume =

  18. [18]

    Large scale simulation of red blood cell aggregation in shear flows , author =. J. Biomech. , volume =

  19. [19]

    The Annals

    Kernel Methods in machine learning , author =. The Annals. of Statistics , volume =

  20. [20]

    Dhairya Malhotra and Abtin Rahimian and Denis Zorin and George Biros , title =

  21. [21]

    Veerapaneni and Denis Zorin and George Biros , title =

    Abtin Rahimian and Shravan K. Veerapaneni and Denis Zorin and George Biros , title =. Journal of Computational Physics , volume = 298, pages =

  22. [22]

    Kelley and David E

    C.T. Kelley and David E. Keyes , title =. SIAM Journal on Numerical Analysis , volume = 35, pages =

  23. [23]

    Mohandas and E

    N. Mohandas and E. Evans , title =. Annu. Rev. Biophys. Biomol. Struct. , volume =

  24. [24]

    Mohandas and P.G

    N. Mohandas and P.G. Gallagher , title =. Blood , volume =

  25. [25]

    Journal of Computational Physics , volume = 347, pages =

    Libin Lu and Abtin Rahimian and Denis Zorin , title =. Journal of Computational Physics , volume = 347, pages =

  26. [26]

    Journal of Computational Physics , volume = 274, pages =

    Bryan Quaife and George Biros , title =. Journal of Computational Physics , volume = 274, pages =

  27. [27]

    Hou and John S

    Thomas Y. Hou and John S. Lowengrub and Michael J. Shelley , title =. 1994 , journal =

    1994

  28. [28]

    Cordasco and A

    D. Cordasco and A. Yazdani and P. Bagchi , title =. Physics of Fluids , volume =

  29. [29]

    Helsing and R

    J. Helsing and R. Ojala , title =. 2008 , journal =

    2008

  30. [30]

    G. R. Baker and M. J. Shelley , title =. 1990 , journal =

    1990

  31. [31]

    Zhang and E-A

    Y. Zhang and E-A. Kim , title =. 2017 , journal =

    2017

  32. [32]

    Snyder and M

    J.C. Snyder and M. Rupp and K. Hansen and K-R. Muller and K. Burke , title =. 2012 , journal =

    2012

  33. [33]

    Wei and R.G

    Q. Wei and R.G. Melko and J.Z.Y. Chen , title =. 2017 , journal =

    2017

  34. [34]

    Wang , title =

    L. Wang , title =. 2016 , journal =

    2016

  35. [35]

    Muravleva and I

    E. Muravleva and I. Oseledets and D. Koroteev , title =. 2018 , journal =

    2018

  36. [36]

    Hornik and M

    K. Hornik and M. Stinchombe and H. White , title =. 1989 , journal =

    1989

  37. [37]

    Chakrabarti and C

    B. Chakrabarti and C. Gaillardac and D. Saintillan , title =. 2020 , journal =

    2020

  38. [38]

    R. T. Fraley and C. S. Fornari and S. Kaplan , title =. 1979 , journal =

    1979

  39. [39]

    LaGrone and R

    J. LaGrone and R. Cortez and W. Yan and L. Fauci , title =. 2019 , journal =

    2019

  40. [40]

    Flormann and O

    D. Flormann and O. Aouane and L. Kaestner and C. Ruloff and C. Misbah and T. Podgorski and C. Wagner , title =. Scientific Reports , volume =

  41. [41]

    S. E. Spagnolie and G. R. Moreno-Flores and D. Bartolo and E. Lauga , title =. 2015 , journal =

    2015

  42. [42]

    Gera and D

    P. Gera and D. Salac and S. E. Spagnolie , title =. 2022 , journal =

    2022

  43. [43]

    Thomas and T

    D. Thomas and T. O'Brien and A. Pandit , title =. 2018 , journal =

    2018

  44. [44]

    Pivkin and Z

    I. Pivkin and Z. Peng and G. E. Karniadakis and P. A. Buffet and M. Dao and S. Suresh , title =. 2016 , journal =

    2016

  45. [45]

    Reichel and J

    F. Reichel and J. Mauer and A. A. Nawaz and G. Gompper and J. Guck and D. A. Fedosov , title =. 2019 , journal =

    2019

  46. [46]

    Henry and S

    E. Henry and S. H. Holm and Z. Zhang and J. P. Beech and J. O. Tegenfeldt and D. A. Fedosov and G. Gompper , title =. 2016 , journal =

    2016

  47. [47]

    G. S. Patterson and S. A. Orszag , title =. 1974 , journal =

    1974

  48. [48]

    Canuto and M

    C. Canuto and M. Y. Hussaini and A. Quarteroni and T. A. Zang , title =. 1987 , address =

    1987

  49. [49]

    Fung , title =

    Y.C. Fung , title =. 1990 , address =

    1990

  50. [50]

    James and D

    G. James and D. Witten and T. Hastie and R. Tibshirani , title =. 2013 , address =

    2013

  51. [51]

    Goodfellow and Y

    I. Goodfellow and Y. Bengio and A. Courville , publisher=. Deep

  52. [52]

    S. A. Orszag , title =. 1971 , journal =

    1971

  53. [53]

    S. A. Orszag , title =. 1974 , journal =

    1974

  54. [54]

    A. G. Kravchenko and P. Moin , title =. 1997 , journal =

    1997

  55. [55]

    2026 , eprint=

    IV-Net: A neural network for elliptic PDEs with random and highly varying coefficients , author=. 2026 , eprint=

    2026

  56. [56]

    Chemical Engineering Journal , volume=

    An improved machine learning approach for predicting granular flows , author=. Chemical Engineering Journal , volume=. 2022 , publisher=

    2022

  57. [57]

    Journal of Machine Learning Research , volume=

    Neural operator: Learning maps between function spaces with applications to pdes , author=. Journal of Machine Learning Research , volume=

  58. [58]

    arXiv preprint arXiv:2411.09678 , year=

    NeuralDEM-Real-time Simulation of Industrial Particulate Flows , author=. arXiv preprint arXiv:2411.09678 , year=

    arXiv

  59. [59]

    npj Biosensing , volume=

    A wearable device for continuous monitoring of circulating cells at single-cell resolution , author=. npj Biosensing , volume=. 2025 , publisher=

    2025

  60. [60]

    Harmon and E

    D. Harmon and E. Vouga and B. Smith and R. Tamstorf and E. Grinspun , title =

  61. [61]

    Computer Methods in Applied Mechanics and Engineering , volume = 200, pages =

    Etienne Vouga and David Harmon and Rasmus Tamstorf and Eitan Grinspun , title =. Computer Methods in Applied Mechanics and Engineering , volume = 200, pages =

  62. [62]

    Besl and Neil D

    Paul J. Besl and Neil D. McKay , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume = 14, pages =

  63. [63]

    Journal of Computational Physics , volume = 306, pages =

    Bryan Quaife and George Biros , title =. Journal of Computational Physics , volume = 306, pages =

  64. [64]

    Dutt and L

    A. Dutt and L. Greengard and V. Rokhlin , title=. BIT Numerical Mathematics , year=

  65. [65]

    Veerapaneni and George Biros , title =

    Abtin Rahimian and Shravan K. Veerapaneni and George Biros , title =. Journal of Computational Physics , year =

  66. [66]

    Journal of Computational Physics , year =

    Lexing Ying and George Biros and Denis Zorin , title =. Journal of Computational Physics , year =

  67. [67]

    Fendler , title =

    Janos H. Fendler , title =. Acc. Chem. Res. , year =

  68. [68]

    J. O. Dabiri and M. Gharib , title =. Proc. R. Soc. B. , year =

  69. [69]

    J. O. Dabiri , title =. Annu. Rev. Fluid Mech. , year =

  70. [70]

    2017 , URL =

    Accelerating Eulerian Fluid Simulation With Convolutional Networks , author =. 2017 , URL =

    2017

  71. [71]

    Power and G

    H. Power and G. Miranda , title =. SIAM Journal of Applied Mathematics , volume =. 1987 , pages =

    1987

  72. [72]

    Power , title =

    H. Power , title =. IMA Journal of Applied Mathematics , volume =

  73. [73]

    Kraus and W

    M. Kraus and W. Wintz and U. Seifert and R. Lipowsky , journal =

  74. [74]

    Ebrahimi and P

    E. Ebrahimi and P. Balogh and P. Bagchi , journal =

  75. [75]

    G. I. Taylor F. R. S. , journal =

  76. [76]

    Lipowsky , journal =

    R. Lipowsky , journal =

  77. [77]

    U. M. Ascher and L. R. Petzold , title =. 1998 , address =

    1998

  78. [78]

    Lipowsky and E

    R. Lipowsky and E. Sackmann , title =. 1995 , address =

    1995

  79. [79]

    B. K. Alpert , title =. SIAM Journal on Scientific Computing , volume =. 1999 , pages =

    1999

  80. [80]

    S. K. Veerapaneni and A. Rahimian and G. Biros and D. Zorin , title =. Journal of Computational Physics , volume =. 2011 , pages =

    2011

Showing first 80 references.