VesNet: Neural network accelerated solver for simulating Stokesian vesicle suspensions
S. Zhong 1, G. Kabacaoğlu 2, G. Biros 1,3 · 2026
Read VesNet: Neural network accelerated solver for simulating Stokesian vesicle suspensions by S. Zhong 1, G. Kabacaoğlu 2, G. Biros 1,3 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25569.txt.
Introduction
Suspensions of soft vesicles, stand ins for red blood cells, are painfully slow to simulate because every membrane interacts with every other through the fluid. S Zhong, G Kabacaoğlu, and G Biros introduce VesNet in 2026, a hybrid solver that neuralises only the expensive local pieces. Networks approximate each vesicle's self interaction, its coupling to the background flow, and the lubrication forces that spike when two vesicles nearly touch. Classical modules still handle membrane reparameterisation and far field interactions.
The graphics processor version ran over a hundred times faster than a leading boundary integral solver on a multicore processor. It also ran about five times faster than the same solver on a graphics processor. Accuracy checks recovered single vesicle shape diagrams, pair dynamics, and thousand vesicle flows in two standard settings. Modest hardware can now run studies that once needed clusters.
Catalog record
| CREATOR | S. Zhong 1, G. Kabacaoğlu 2, G. Biros 1,3 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 79 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.25569
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
VesNet: Neural network accelerated solver for simulating Stokesian vesicle suspensions S. Zhong 1, G. Kabacaoğlu 2, G. Biros 1,3 Abstract Numerically simulating suspensions of deformable particles in the Stokes flow regime is computationally demanding due to strongly nonlinear fluid–structure interactions, evolving interfaces, and multiscale hydrodynamic phenomena. We introduce VesNet, a framework developed to significantly accelerate simulations of vesicle suspensions in two dimensions. Vesicles are deformable particles that maintain local surface inextensibility and resist bending. VesNet comprises components that approximate vesicle self-interactions, including coupling to the imposed background flow and short-range lubrication forces that arise in near-contact situations. Instead of acting as a purely neural surrogate model, VesNet is a hybrid approach that also retains conventional algorithmic modules for boundary reparameterization and N-body far-field hydrodynamic interactions. A GPU-accelerated implementation of VesNet delivers more than a 100 × \times speedup over a multithreaded MATLAB CPU implementation of a state-of-the-art, fully resolved boundary integral equation (BIE) solver, and roughly a 5 × \times speedup relative to a GPU implementation of the same BIE solver. To evaluate its fidelity, we analyze the recovery of single-vesicle phase diagrams, the dynamics of interacting vesicle pairs, and large-scale simulations with thousands of vesicles in Taylor–Green and Poiseuille flow settings. VesNet successfully reproduces the main observables of interest, thereby enabling extensive simulations using relatively modest computational resources. † † affiliation: 1 Oden Institute for Computational Science and Engineering, The University of Texas at Austin, Austin, TX-78712, USA 2 Department of Computer Science, Durham University, Durham, DH1 3LE, UK 3 Walker Department of Mechanical Engineering, The University of Texas at Austin, Austin, TX-78712, USA 1 Introduction Particulate suspensions play a vital role in diverse domains, from biological systems ( 12 ) and materials science ( 10 ) , to advanced manufacturing ( 19 ) , and soft robotics ( 2 ) . These systems often operate at low Reynolds numbers, where viscous forces dominate, placing their dynamics within the Stokes flow regime. Modeling such systems is challenging due to the complex, multiscale hydrodynamic interactions, which exhibit long-range many-body effects and short-range lubrication forces ( 52 ; 55 ; 1 ; 16 ; 41 ; 62 ) . We demonstrate VesNet’s computational accuracy and performance on a specific particulate flow: vesicle flows (see Fig. 1 ). Vesicles are highly deformable yet inextensible membranes containing a Newtonian fluid ( 61 ) . They resist stretching, compression, and bending, and have nearly constant surface area and enclosed fluid volume. Vesicles serve as numerical and experimental proxies of red blood cells ( 14 ; 42 ; 22 ; 57 ) . Fast and accurate simulations of vesicle suspensions provide an indispensable tool to discover optimal designs of microfluidic devices in medical applications and beyond ( 20 ; 33 ) . Figure 1 : 2,000 vesicles in Poiseuille flow (see sections 5.3 and 5.5 for details). The characteristic flow time scale is t_{\dot{\gamma}}=1/\dot{\gamma} , with \dot{\gamma}=u_{\max}/W , where u_{\max} denotes the peak velocity at the channel center and W is the channel width. The initial vesicle area fraction is 30%. Under the imposed flow, vesicles disperse and interact with each other. Each row shows the suspension at successive times. The left column displays the full suspension, while the right column provides magnified views of the regions enclosed by the black rectangular windows in the left column. VesNet combines several neural networks for self-interactions, advection, and near-singular evaluation, together with N-body interactions and spectral operators for computing elastic forces, performing interpolation, and enforcing area and length conservation. The complete simulation is executed on a GPU for 4000 steps with a time step of \Delta t=10^{-5} . A single VesNet time step for this simulation takes 1.4 s on one NVIDIA GH200. 1.1 Contributions VesNet stands out distinctly from existing particulate flow simulation methods based on machine learning (ML) due to its integration of neural networks (NN) with the boundary integral equation method (BIEM) and its focus on deformable particles. Building on MLARM ( 21 ) , we enhance both accuracy and performance to enable dense vesicle suspension simulations with highly deformable vesicles. The resulting ML-accelerated solver, VesNet, features: • VesNet approximates the integro-differential operators using the Iterated V-shape Net (IVNet) introduced by ( 71 ) . IVNet captures the fact that the solution to the Stokesian particulate flows involves convolutions with the fundamental solution to the Stokes equations and has a multiscale nature. This network architecture enables accurate approximations of the nonlinear operators. • VesNet uses an an IVNet to handle nearly-singular integrals for vesicles in close contact to enable dense suspension simulations and introduces an algorithm to efficiently couple it to arbitrary suspensions. • VesNet is fully GPU accelerated; all steps, including solver steps and the network inference run on a GPU. This enables fast simulations of vesicle flows, offering a robust tool for scenarios such as parameter exploration, optimization problems, and uncertainty quantification where traditional simulations and existing ML methods fall short. The code is open source and released at https://github.com/bioflume/Ves2Dpy . VesNet embeds physical invariances—such as translation, rotation, and scaling—directly into both its training data and architecture, resulting in a generalizable and interpretable model. This design enables VesNet to achieve BIEM-level accuracy at reduced spatial and temporal resolutions. 1.2 Related work High fidelity methods for suspensions of deformable particles. Simulation methods for particulate flows typically fall into two categories. First, methods based on solvers for boundary value problems, such as BIEM ( 64 ) and immersed boundary methods ( 46 ) , require solving large linear systems at every time step, making them computationally expensive for large particle counts. Second, particle methods, including Stokesian Dynamics ( 9 ) , Smoothed Particle Hydrodynamics ( 44 ) , and Dissipative Particle Dynamics ( 48 ; 8 ) , bypass PDE solutions but still involve intensive computations like matrix inversions, especially for large-scale systems. Of course, this list is not exhaustive, and other methods (e.g., Lattice Boltzmann) and hybrid versions exist. In this study, we present VesNet , a ML-accelerated solver that synergistically integrates ML algorithms with BIEM to address the inherent computational bottlenecks of conventional BIEM frameworks. This hybrid approach enables rapid and precise simulations of two-dimensional (2D) dense particulate Stokes flows. The methodology developed here establishes a foundational framework for future extension to 3D systems, paving the way for high-performance simulations at previously unattainable scales. Machine learning algorithms for suspensions. Recent advances in ML have opened transformative avenues for modeling complex physical systems, particularly those involving interacting particles. ML is used for approximating force fields ( 70 ) , deriving coarse-grained models ( 39 ) , approximating hydrodynamic interactions ( 47 ) and enabling multiscale modeling ( 66 ) . Most particulate flow simulations using ML consider rigid particles. While a vast body of research exists on this topic, we refer to the recent studies to highlight the trends ( 5 ; 6 ) . One of the most recent studies is NeuralDEM 3 . It replaces traditional, computationally intensive discrete element method (DEM) simulations with fast, adaptable ML surrogates. This model captures long-term transport processes across various regimes using macroscopic observables, eliminating the need for detailed microscopic parameters. Another study for DEM is by 67 where ML is integrated into DEM through