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

TL;DR

This paper presents a deep graph neural network approach to accurately predict the frictional contact network in dense suspensions near jamming, overcoming experimental and computational challenges.

Contribution

It introduces a novel GNN-based method that generalizes well across diverse conditions to predict contact networks in suspensions close to jamming.

Findings

Accurately predicts FCNs across wide phase space

Demonstrates robust generalization and extrapolation

Applicable to various suspension states and parameters

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 (${\sigma}_{xy}$), packing fraction(${\phi}$) and sliding and rolling friction (${{\mu}_s, {\mu}_r}$). 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.