Paper Number
PO31 My Program
Session
Poster Session
Title
Extrapolation and interpolation of force chain networks in dense suspensions employing graph neural network
Presentation Date and Time
October 16, 2024 (Wednesday) 6:30
Track / Room
Poster Session / Waterloo 3 & 4
Authors
- Aminimajd, Armin (Case Western Reserve University, Macromolecular Science and Engineering)
- Maia, Joao (Case Western Reserve University, Macromolecular Science and Engineering)
- Singh, Abhinendra (Case Western Reserve University, Macromolecular Science and Engineering)
Author and Affiliation Lines
Armin Aminimajd, Joao Maia and Abhinendra Singh
Macromolecular Science and Engineering, Case Western Reserve University, Cleveland, OH 44106
Speaker / Presenter
Aminimajd, Armin
Keywords
computational methods; data-driven rheology; dense systems; future of rheology; non-Newtonian fluids; techniques
Text of Abstract
The Force Chain Network (FCN), originating from static and rolling frictional contacts, is believed to be linked to the shear thickening mechanism. The stress-driven transition from lubricated unconstrained to constrained pairwise particle motion leads to complex non-Newtonian behaviors in dense suspensions. Lubrication Flow Discrete Element Modeling (LF-DEM) has successfully captured the complex rheological behaviors of dense suspensions and mesoscale FCNs. However, traditional simulation methods are costly and resource-intensive, and experimental techniques are often inaccurate or limited to specific systems. In this study, we predict the structure and occurrence of FCNs using a Graph Neural Network (GNN). By employing this method, we convert the datasets generated from LF-DEM into graphs, representing particles and interparticle interactions as nodes and edges, respectively. After training, we demonstrate the scalability and robustness of our model in extrapolating and interpolating FCNs across various system particle sizes, packing fractions, shear stresses, particle stiffnesses, and different surface effects, including sliding and rolling friction constants. Our model achieves accuracies exceeding 98%, even when trained on data far from its control parameters. This GNN technique can serve as a powerful tool for predicting rheological behaviors and characterizing complex particulate systems.