Scalability of a graph neural network in accurate prediction of frictional contact networks in suspensions

Abstract

Dense suspensions often exhibit shear thickening, characterized by a dramatic increase in viscosity under large external forcing. This behavior has recently been linked to the formation of a system-spanning frictional contact network (FCN), which contributes to increased resistance during deformation. However, identifying these frictional contacts poses experimental challenges and is computationally expensive. This study introduces a graph neural network (GNN) model designed to accurately predict FCNs by two dimensional simulations of dense shear thickening suspensions. The results demonstrate the robustness and scalability of the GNN model across various stress levels (σ), packing fractions (ϕ), system sizes, particle size ratios (Δ), and amounts of smaller particles. The model is further able to predict both the occurrence and structure of a FCN. The presented model is accurate and interpolates and extrapolates to conditions far from its control parameters. This machine learning approach provides an accurate, lower cost, and faster predictions of suspension properties compared to conventional methods, while it is trained using only small systems. Ultimately, the findings in this study pave the way for predicting frictional contact networks in real-life large-scale polydisperse suspensions, for which theoretical models are largely limited owing to computational challenges.