(629f) Application of Graph Neural Network in Accurate Prediction of Force Chain Network in Dense Suspensions

Concentrated or dense, packings of particulate systems in a liquid or air are commonly found in natural, human health, and industrial settings [1]. These systems range from blood to paint, concrete, etc. They are often out-of-equilibrium and display a wide array of rheological features, of which the liquid–solid phase transition under external deformation is arguably the most interesting. Over the last decade, research activity at the intersection of fluid mechanics, granular materials, driven disordered systems, and soft condensed matter physics has led to a consensus that the formation of mesoscale stress-activated frictional contact network (FCN) drives these mechanisms [2,3]. Particle simulations that led to this concept have been successful in quantitatively reproducing the non-Newtonian behavior of thickening suspensions [4]. Inferring when and where these FCNs appear is of vital importance in physics, geology, engineering, and industry due to their impact on material properties. While traditional simulation methods are expensive and time-consuming, recent deep learning techniques were found to be a powerful tool to simulate and predict properties of dry granular materials, colloidal, and complex physical systems. Among these, graph neural network (GNN) has proven to be robust and fast, using nodes and edges to represent particles and their relationships. Herein, we train the deep graph convolutional neural network (GCNN) model using datasets obtained from lubrication flow discrete element modeling (LF-DEM) to accurately predict FCN in suspensions at different volume fractions, step strains, stresses, disorder degrees, stiffness coefficient and friction coefficient. Our machine learning model goes beyond recent research on prediction of contact forces in granular materials; it not only predicts FCN with higher accuracy but also interpolates and extrapolates to conditions far from control parameters. The method used in this study can be utilized for prediction of rheological and characterization of complex particulate systems in the future. Once extensive validation is performed, these systems have capabilities to speed up current simulation and continuum modeling efforts, which can contribute to extraterrestrial building efforts.

References:

1. Morris, J. F. Shear Thickening of Concentrated Suspensions: Recent Developments and Relation to Other Phenomena. Annu. Rev. Fluid Mech. 52, 121–144 (2020).
2. Singh, A. Hidden hierarchy in the rheology of dense suspensions. MRS Communications. 13 (6), 971-979 (2023).
3. M van der Naald, A Singh, TT Eid, K Tang, JJ de Pablo, HM Jaeger. Minimally rigid clusters in dense suspension flow, Nature Physics, 1-7, 2024.
4. Singh, A., Jackson, G. L., van der Naald, M., de Pablo, J. J. & Jaeger, H. M. Stress-activated Constraints in Dense Suspension Rheology. Physical Review Fluids 7 (5), 054302 (2022).