Prediction of Frictional Contact Networks Using Deep Graph Convolutional Neural Network in Dense Suspensions. Part 1: Methods and Scalability

The shear thickening behavior in dense suspensions has recently been linked to a shift from lubrication-dominated state, where suspension flows, to a state characterized mainly by constrained interparticle interactions. Here, the constraints are considered to be originated from frictional contacts between particles. Under external deformations, particles in frictional contact restrict the suspension flow and hence result in the increase of viscosity. In this work, we predict the structure of the frictional contact network (FCN) at various conditions. For the simulation of suspensions, lubrication flow discrete element modeling (LF-DEM) has been employed successfully to quantitatively capture the non-Newtonian shear rheology of dense suspensions. While valuable, traditional simulation techniques are time-demanding and require huge energy resources when simulating dense suspensions. Recent deep learning techniques have emerged as an efficient tool for the prediction of particulate system properties, proven to outperform conventional simulation methods. In this study, DeepGCN (Deep Graph Convolutional Network), a variant of GNN (Graph Neural Network) has been utilized to train models to predict FCN in suspensions with different scales, shear stresses, particle size ratios, and volumetric mixing ratios. Our model can interpolate and extrapolate FCN far from its initial training conditions. This technique has the ability to characterize complex soft matter systems in the future.