Multitask Prediction of Rheological Properties in Suspensions Using Graph Neural Network

Efficient prediction of suspension rheology is critical for optimizing processing and performance across diverse industries. Traditional experimental approaches face challenges related to reproducibility and scalability, while computational methods are often resource-intensive and time-consuming. This study introduces a multitask learning framework utilizing Graph Neural Networks (GNNs) to predict bulk properties of suspensions—including particle pressure, viscosity, and friction coordination—Based on physical intuition. Simulations of two-dimensional suspensions spanning semi-dilute to dense concentrations were conducted to capture a wide spectrum of non-Newtonian behaviors, from continuous and discontinuous shear thickening to jamming, using Lubrication-Flow Discrete Element Modeling (LF-DEM). GNN models were independently trained across a wide range of packing fractions (φ = 0.70 to 0.80) and with shear stress values. The prediction results showed strong correlation coefficients (R² = 0.99) and minimal mean absolute error for packing fractions up to φ = 0.78 while a slight decrease observed at φ = 0.80 likely due to fluctuations near jamming condition. Nevertheless, the fitting curves indicates significant agreement between simulated and predicted outcomes. Finally, this GNN-based multitask learning framework proves to be a promising and efficient method for predicting the properties of particulate systems under diverse conditions.