Development of a XGBoost-based drag force model for freely evolving particle suspensions

Abstract

An XGBoost-based drag model is developed using data from Particle Resolved Simulations (PRS) of freely evolving spherical particle suspensions, encompassing Reynolds numbers from 10 to 300, solid volume fraction between 0.1 and 0.4, and particle-to-fluid density ratio of 2, 10 and 100. Drag force data from 150 continuous time instances in PRS are divided into two sets: the first set that includes data from the initial 120 instances is used for training the model and interpolation testing, while the second set comprises drag forces from the final 30 instances is used exclusively for extrapolation testing. Both interpolation and extrapolation tests demonstrate significantly improved accuracy compared to traditional drag correlations. Notably, the model achieves its highest prediction accuracy for particles with density ratios of 100, which is attributed to the increased influence of unsteady drag forces at lower density ratios that cannot be fully captured by instantaneous particle distributions alone.