Learning and Interpreting Drag Force Models for Dense Particle Suspensions with Graph Neural Networks

Neil Ashwin Raj, Nikhil Muralidhar, Ze Cao, Danesh Tafti

Abstract

Fluid flow across random particle arrangements appears in various natural and engineering systems. Experimental studies on these systems can be challenging due to the need for costly equipment. Numerical methods offer an alternative but require a trade-off between computational cost and accuracy. Particle Resolved Simulations (PRS) provide high accuracy but are time-consuming, while CFD-DEM methods are more efficient but less precise, as they estimate a mean drag force for particle groups instead of individual particles. Previous studies have used deep learning for drag force predictions in particle suspensions, yet few have explored graph-based models. In this study, we introduce a curriculum learning framework, outperforming traditional graph training methods. Additionally, our model is trained on data from non-stationary particle simulations, unlike prior supervised drag models. Including particle mobility in our models further improves predictions, especially at higher Reynolds numbers and solid fractions.