Robust Prediction of Frictional Contact Network in Near-Jamming Suspensions Employing Deep Graph Neural Networks

Armin Aminimajd, Joao Maia, Abhinendra Singh · 2025

Read Robust Prediction of Frictional Contact Network in Near-Jamming Suspensions Employing Deep Graph Neural Networks by Armin Aminimajd, Joao Maia, Abhinendra Singh (2025) 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-2502-18743.txt.

Introduction

Armin Aminimajd, Joao Maia, and Abhinendra Singh present a 2025 paper on predicting the force network inside thick suspensions. A dense suspension of grains in liquid stiffens sharply near jamming, when grains lock into a load-bearing web of frictional contacts. Measuring or simulating that web near jamming is hard and slow.

A deep graph neural network learned the web from simulation data, treating grains as nodes and contacts as links. Trained on a single flow condition, it still predicted networks across wide ranges of density and flow, from dilute to jammed and from startup to steady flow. The approach offers a cheap stand-in for costly simulations of pastes, slurries, and other dense mixtures.

Catalog record

CREATORArmin Aminimajd, Joao Maia, Abhinendra Singh
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-18743

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Robust Prediction of Frictional Contact Network in Near-Jamming Suspensions Employing Deep Graph Neural Networks Armin Aminimajd, Joao Maia, Abhinendra Singh Abstract The viscosity of the suspension consisting of fine particles dispersed in a Newtonian liquid diverges close to the jamming packing fraction. The contact microstructure in suspensions governs this macroscopic behavior in the vicinity of jamming through a frictional contact network (FCN). FCN is composed of mechanical load-bearing contacts that lead to the emergence of rigidity near the jamming transition. The stress transmission and network topology, in turn, depend sensitively on constraints on the relative motion of the particles. Despite their significance, predicting the FCN, especially close to jamming conditions, remains challenging due to experimental and computational impediments. This study introduces a cost-effective machine learning approach to predict the FCN using a graph neural network (GNN), which inherently captures hidden features and underlying patterns in dense suspension by mapping interparticle interactions. Employing a variation of GNN called the Deep Graph Convolutional Network (DeepGCN) trained on data-driven simulations, this study demonstrates robust generalization and extrapolation capabilities, accurately predicting FCNs in systems with divergent flow parameters and phase spaces, despite each being trained exclusively on a single condition. The study covers a wide range of phase space, from semi-dilute to jammed states, spanning transient to steady states, while systematically varying parameters such as shear stress ( σ xy \sigma_{\text{xy}} ), packing fraction ( ϕ \phi ) and sliding and rolling friction ({ μ s , μ r \mu_{s},\mu_{r} }). The results of this research pave the way for innovative transferable techniques in predicting the properties of particulate systems, offering new avenues for advancement in material science and related fields. I Introduction Particulate suspensions, consisting of solid particles dispersed in a liquid medium, are ubiquitous in natural and industrial contexts. Examples range from everyday materials such as paints, concrete, food products, and pharmaceuticals to complex natural systems, including blood flow, mudflow, and lava Coussot 1997 ; Jerolmack and Daniels 2019 ; Galdi et al. 2008 ; Javid and Toufigh 2024 . During the last decade, a unifying framework has emerged, linking the physics of dense non-Brownian suspensions with dry granular rheology Boyer, Guazzelli, and Pouliquen 2011 ; Guazzelli and Pouliquen 2018 ; Morris 2020 ; Singh 2023 . The key insight is that, under large deformations, the lubrication film between non-Brownian particles ruptures, leading to the formation of direct frictional contacts that share similarities with dry granular materials Mari et al. 2014 ; Singh et al. 2018 ; Singh et al. 2020 ; Morris 2020 ; Ness, Seto, and Mari 2022 ; Guy, Hermes, and Poon 2015 ; Wyart and Cates 2014 . Further studies have extended this understanding to the emergence of a frictional contact network (FCN) in dense suspensions, drawing parallels to classical two-dimensional granular material experiments by Behringer and coworkers Bi et al. 2011 ; Majmudar and Behringer 2005 ; Majmudar et al. 2007 ; Clark, Kondic, and Behringer 2012 ; Tordesillas, Zhang, and Behringer 2009 ; Daniels and Hayman 2008 ; Daniels, Kollmer, and Puckett 2017 . Figure 1: Schematic for the simulation process and the prediction of Frictional Contact Network (FCN) Using Deep Graph Convolutional Neural Network (DeepGCN). Initially, configurations are generated by the LF-DEM method, where particles obey Stokes flow (Eq. ( 1 )). Then the configurations including the information of all particles (either in frictional contacts or not) are transformed into graphs readable for GNN where particles and the interparticle interactions are represented by nodes and edges, respectively. The graphs contain particle radius and relative distance between particles as node feature ( h_{u} ) and edge attributes ( e_{v}u ), respectively. Next, they are processed into residual graph convolution layers where nodes are updated through the message passing process using information of neighborhood nodes and edges. After normalization, a non-linear activation function (ReLU) applies to improve the model’s understanding of complex underlying particle relationships. Then, a final linear layer classifies nodes to a specific class, whether a particle participates in the formation of FCN or not. Misclassified particles by the model, either incorrectly identified as part of the FCN or excluded when they should be included, are highlighted in red. The suspension of interest in this study ranges from semi-dilute to dense systems near jamming. As the volume fraction \phi increases, the viscosity increases, eventually diverging as volume fraction \phi approaches the frictional jamming limit \phi_{J}^{\mu} Peters, Majumdar, and Jaeger 2016 ; Barik and Majumdar 2022 ; Singh et al. 2019 ; Ness and Moussa 2024 . The frictional jamming point \phi_{J}^{\mu} is influenced by various particle surface properties, including roughness, shape, interfacial chemistry, and particle size distribution, among other factors Morris 2020 ; Singh et al. 2020 ; Singh et al. 2022 ; Singh 2023 ; Singh et al. 2024 ; James et al. 2018 ; Hsiao and Pradeep 2019 ; Hsu et al. 2018 . Recognizing the critical role of \phi_{J}^{\mu} in dense suspension rheology, numerous experimental studies have explored ways to manipulate surface morphology Hsiao et al. 2017 ; Hsiao and Pradeep 2019 ; Pradeep et al. 2021 ; Hsu et al. 2018 and interfacial chemistry Sivadasan et al. 2019 ; James et al. 2018 ; James et al. 2019 ; Chen et al. 2023 ; van der Naald et al. 2021 ; Xu, Singh, and Jaeger 2020 ; Lee et al. 2021 . Fundamentally, these surface modifications alter the “effective friction” by imposing constraints on the relative motion between particles, including translational, rotational, and twisting modes Singh et al. 2020 ; Guy et al. 2018 ; Singh et al. 2022 ; Singh 2023 . Figure 2: Model Training Process for FCN Prediction. An illustrative example of a model trained at \phi=0.7 and \sigma/\sigma_{0}=5 , using a dataset of 320 graphs for training and 80 graphs for validation. As the loss value decreases over epochs, the prediction accuracy correspondingly improves. The training has been implemented using only two hidden layers with hidden dimensions set to 64. The process stops using the early stop technique that prevents overfitting. The picture that emerges is that the constraints on relative motion stabilize the load-bearing force network, thereby enhancing the correlated motion Singh et al. 2020 ; Liu and Nagel 2010 . In addition to that, these particle-level contacts give rise to large-scale force and contact networks that resist large deformations Jamali, Del Gado, and Morris 2020 . Recently, network science tools have been employed to characterize the statistical properties of frictional force and contact networks and to correlate them with the rheological behavior of suspensions Thomas et al. 2018 ; E. Thomas et al. 2020 ; Nabizadeh, Singh, and Jamali 2022 ; Sedes et al. 2022 ; Edens et al. 2019 ; Edens et al. 2021 ; Gameiro et al. 2020 ; van der Naald et al. 2024 ; D’Amico, Tu, and Singh 2025 . Studies on dry granular materials and dense suspensions Thomas et al. 2018 ; E. Thomas et al. 2020 ; Blumenfeld and Edwards 2003 ; Blumenfeld and Edwards 2003 further suggest that the network space orthogonal to particle space may hold the key to establishing a statistical mechanics framework for out-of-equilibrium systems. Given this, significant efforts have been made to deduce interparticle

Read the complete text