A supervised machine learning approach for predicting variable drag forces on spherical particles in suspension

CFD-DEM simulations have been used extensively to study dense fluid-particle systems. In the point mass representation of particles in DEM, the modeled drag force plays an important role in the dynamics. Current state-of-the-art methodologies use the mean drag correlations based on the superficial Reynolds number and void fraction. In this work, as proof-of-concept, a new data-driven approach for drag force model development is presented using Particle Resolved Simulations (PRS). The key idea in the proposed framework is the use of supervised machine learning to build higher fidelity drag force models for CFD-DEM simulation based on data obtained by PRS. Results show that a trained artificial neural network (ANN) improves the accuracy of drag force prediction by accounting for the relative neighbor particle locations as inputs to the model along with the existing Reynolds number and void fraction information. The ANN trained prediction are within 15% of PRS predictions for 68% of particles, whereas only 46% of particles lie in the same error range if the mean drag is used. This work highlights the viability and potential of using machine learning to develop accurate drag models for particles in suspension.