Inferring the Turbulent Breakup of Colloidal Aggregates Using Graph Neural Networks Journal: International Journal of Multiphase Flow
M. Buzzicotti, M. Cencini, G. Cimini, M. Vanni, A. S. Lanotte · 2026
Read Inferring the Turbulent Breakup of Colloidal Aggregates Using Graph Neural Networks Journal: International Journal of Multiphase Flow by M. Buzzicotti, M. Cencini, G. Cimini, M. Vanni, A. S. Lanotte (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-2601-05667.txt.
Introduction
Michele Buzzicotti and 4 co-authors use graph neural networks, in a 2026 paper, to predict whether colloidal aggregates will break apart in turbulent flow. Solid clumps suspended in a fluid can fragment when shear stresses pull at them. Knowing the aggregate's structure and the surrounding velocity gradients may be enough to forecast that breakup.
One network classifies aggregates into those that break and those that survive. A second predicts the largest tensile force inside each aggregate under given flow settings. Both performed with high statistical accuracy and general to aggregates of different sizes, beating predictions based on averaged quantities. The approach could help quantify breakup across large populations in powder production and sediment transport.
Catalog record
| CREATOR | M. Buzzicotti, M. Cencini, G. Cimini, M. Vanni, A. S. Lanotte |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-05667
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Inferring the Turbulent Breakup of Colloidal Aggregates Using Graph Neural Networks Journal: International Journal of Multiphase Flow M. Buzzicotti, M. Cencini, G. Cimini, M. Vanni, A. S. Lanotte Abstract Solid aggregates in turbulent suspensions may break under the action of shear stresses. We explore the use of Graph Neural Networks (GNN) to infer aggregate fragmentation once the aggregate structure and flow velocity gradients are known. We consider two models: the first GNN is a classifier, trained to distinguish aggregates that break from those that do not; the second GNN is a regression model, trained to predict the maximal tensile force within each aggregate in a given flow condition. We show that both models complete their task with a high statistical accuracy, and generally perform better than the statistical prediction based on mean field quantities. This work paves the way for future use of Graph Neural Networks to quantify aggregate breakup in large population of aggregates suspended in complex flow configurations. 1 Introduction In turbulent suspensions, the breakup of agglomerates – material structures formed from an ensemble of fragments or primary particles or monomers compacted together – due to the hydrodynamic stress generated by the surrounding fluid is a key process in a wide range of applications, in both natural and engineering contexts. Examples include rubber compounding ( 32 ) , the production of micro-particles via wet synthesis ( 18 ) , the rheological behavior of suspensions ( 5 ) , and sediment transport in natural flows or wastewater streams ( 26 ) . Despite the broad scope of these applications, only a few experimental studies have analyzed the breakup process at the scale of individual agglomerates ( 21 ; 39 ; 38 ; 34 ) . On the numerical side, most of the approaches (based, e.g., on Discrete Element Methods or Immersed Boundary methods) for the turbulent breakup of aggregates demand complex implementation and considerable computational resources. As a consequence, their application soon becomes impractical for large populations of agglomerates when directly coupled with complex flow fields obtained by integrating the three-dimensional Navier-Stokes equations. An alternative strategy, which we explore here, is to develop surrogate models based on neural networks (NN), capable to determining the breakup condition of an agglomerate after appropriate training. These models are generally computationally fast and could enable simulations involving very large aggregate populations, thereby providing a realistic representation of process equipment, thus representing a promising avenue of numerical modeling. Recently, 27 trained a NN on Large Eddy Simulations of a turbulent flow coupled with Discrete Element Methods to quantify breakup due to inter-agglomerate collisions; the trained model was then integrated into an Eulerian–Lagrangian simulation framework, replacing the physical collision model with a much faster neural surrogate. A model extension included the agglomeration process also ( 28 ) . In this study, we present a data-driven approach, based on Graph Neural Networks (GNNs), which is meant to make predictions on aggregate breakup without performing simulations to directly compute the forces acting on the aggregate. In particular, we develop two different GNN models: the first one is simply meant to classify whether a single aggregate will break or not; the second is meant to infer the maximal tensile force acting on one of its internal links. Using the latter model, the inferred maximal force can then be compared to a threshold value for rupture to establish whether the aggregate will break or not. We focus on small agglomerates composed of few hundreds of primary particles, and whose breakup is induced by the hydrodynamic stresses generated by the turbulent flow only ( 29 ; 9 ) , rather than by the combination of different processes (flow stresses, inter-agglomerate collisions or impacts with the walls). This study is thus an attempt to develop GNNs to study aggregates breakup on three-dimensional turbulent suspensions, a complex system characterized by both large-scale coherent structures and highly non-Gaussian and intermittent small-scale fluctuations. In the last few years, Graph Neural Networks ( 45 ; 19 ; 7 ) have become a standard tool for machine learning on graph-structured data, that is, data represented as a set of entities (nodes) with features, and of a set of relationships (links) among them. Initially mainly used to solve node classification and link inference tasks ( 55 ) , nowadays GNNs have a wide span of applications across domains, including chemistry and material science ( 35 ; 49 ; 8 ) . In fluid dynamics, GNNs have been used for different goals. Since they operate on graphs, which can be converted from any mesh, GNNs can ensure a significantly better resolution for relevant regions within the domain with the same mesh size (see e.g. 2 ), compared with standard convolutional neural networks. While GNNs have been used to simulate the complex fluid dynamical problems under a wide range of conditions ( 40 ; 47 ; 31 ; 17 ; 13 ) , their application to solid aggregates in turbulent suspensions remains a relatively unexplored research area. The paper is organized as follows. In Sec. 2 , we introduce the system under consideration and the numerical modeling of the turbulent suspension. Then, in Sec. 3 , we details the databases (DBs) used to train and test the GNNs. The Graph Neural Network models - classifier and regression- are introduced in Sec. 4 , while a statistical test to be compared to the Neural Networks results is presented in Sec. 5 . Section 6 is devoted to discussing the performances of the two GNNs on different datasets, and the origin of model failures. We summarise our work in Section 7 . 2 Numerical modeling of turbulent suspensions The simplest numerical methods to study aggregate breakup use the free-draining approximation, where each primary particle is assumed to be unaffected by the flow perturbations caused by neighboring particles ( 3 ; 11 ; 46 ; 37 ) . Estimates of the screening effect of hydrodynamic forces within agglomerates have been based on the fraction of the primary particle surface area directly exposed to the external fluid flow ( 23 ; 12 ) . More sophisticated semi-empirical methods, that relate screening to the solid volume fraction or aggregate geometry, were later adopted by 53 and 54 . These approaches fall within the Eulerian–Lagrangian framework, in which aggregates are treated as point-like entities. A more computationally demanding alternative involves full resolution of the fluid flow at the scale of the primary particles. This route was taken by different authors by means of Lattice-Boltzmann simulations to investigate the restructuring and breakup of agglomerates in shear ( 42 ; 43 ) , accelerated ( 44 ) and turbulent flows ( 10 ) . For colloidal aggregates with very small particle Reynolds number, Re_{p}=\gamma_{eff}R^{2}/\nu\ll 1 ( R being the aggregate size in terms of its radius of gyration, \nu the fluid kinematic viscosity and \gamma_{eff} the effective shear rate), Stokesian Dynamics (SD) offers an equally accurate alternative, enabling precise computation of the hydrodynamic forces acting on the aggregates. Compared to techniques that fully resolve the solid–liquid interfaces, and the hydrodynamic forces/torques acting on each primary particle, SD is less computationally expensive and allows for the study of systems with several hundreds of primary particles, in contrast to