Deep Learning Methods for Predicting Fluid Forces in Dense Particle Suspensions

Modelling solid-fluid multiphase flows are crucial to many applications such as fluidized beds, pyrolysis and gasification, catalytic cracking etc. Accurate modelling of the fluid-particle forces is essential for lab-scale and industry-scale simulations. Fluid-particle system solutions can be obtained using various techniques including the macro-scale TFM (Two fluid model), the meso-scale CFD-DEM (CFD - Discrete Element Method) and the micro-scale PRS (Particle Resolved Simulation method). As the simulation scale decreases, accuracy increases but with an exponential increase in computational time. Since fluid forces have a large impact on the dynamics of the system, this study trains deep learning models using micro-scale PRS data to predict drag forces on ellipsoidal particle suspensions to be applied to meso-scale and macro-scale models. Two different deep learning methodologies are employed, multi-layer perceptrons (MLP) and 3D convolutional neural networks (CNNs). The former trains on the mean characteristics of the suspension including the Reynolds number of the mean flow, the solid fraction of the suspension, particle shape or aspect ratio and inclination to the mean flow direction, while the latter trains on the 3D spatial characterization of the immediate neighborhood of each particle in addition to the data provided to the MLP. The trained models are analyzed and compared on their ability to predict three different drag force values, the suspension mean drag which is the mean drag for all the particles in a given suspension, the mean orientation drag which is the mean drag of all particles at specific orientations to the mean flow, and finally the individual particle drag. Additionally, the trained models are also compared on their ability to test on data sets that are excluded/hidden during the training phase. For instance, the deep learning models are trained on drag force data at only a few values of Reynolds numbers and tested on an unseen value of Reynolds numbers. The ability of the trained models to perform extrapolations over Reynolds number, solid fraction, and particle shape to predict drag forces is presented. The results show that the CNN performs significantly better compared to the MLP in terms of predicting both suspensions mean drag force and also mean orientation drag force, except a particular case of extrapolation where the MLP does better. With regards to predicting drag force on individual particles in the suspension the CNN performs very well when extrapolated to unseen cases and experiments and performs reasonably well when extrapolating to unseen Reynolds numbers and solid fractions.