A Multi-Task Model Based on Graph Attention Networks for the Simulation of Two-Phase Flow in Concentrated Suspensions

This paper proposes a multi-task model based on Graph Attention Networks (GATs) for simulating two-phase flow in concentrated suspensions. Concentrated suspensions are widely used in both natural and industrial flows and exhibit complex rheological behaviors such as flow-induced migration, shear thinning, and shear thickening, which significantly impact production processes. Traditional Computational Fluid Dynamics (CFD) models, while effective in multiphase flow modeling, often incur high computational costs when dealing with complex systems. Therefore, we explore data-driven models based on GATs as an alternative to mechanistic models. Using simulation data generated by COMSOL Multiphysics, we construct graph-structured data and apply the multi-task GAT model for prediction. Experimental results show that the model achieves lower Root Mean Square Error (RMSE) and higher coefficient of determination (R²) compared to Long Short-Term Memory (LSTM) and Graph Convolutional Network (GCN) models when predicting parameters such as solid volume fraction, velocity, and pressure. This demonstrates that the GAT-based multi-task model can effectively capture the spatial and temporal characteristics of two-phase flow, offering an efficient and accurate approach for modeling dense suspension two-phase flow.