Meshfree dispersed multiphase flow prediction using equivariant neural network

Deterministic influence of neighboring particles on a particle of interest is not accounted for in a typical Euler-Lagrange simulation.Recent efforts have shown that Deep Learning methods hold promise in capturing these influences.The considered problem of flow around a particle in a unique neighbor configuration involves rotational symmetry about the mean-flow direction.However, conventional neural networks do not implicitly satisfy symmetries.Thus, this work makes use of equivariant neural networks that inherently preserve rotational symmetry.We believe that accurate flow prediction is a crucial intermediate step to eventually obtain robust particle forces. Achieving flow predictions using a Convolutional Neural Network at the resolution of Particle-Resolved Direct Numerical Simulation is computationally intensive.Hence, a meshfree flow method based on point-cloud equivariant neural network to obtain flow field predictions is proposed.This computationally efficient methodology produces results that are 75-90% accurate for particle volume fraction and Reynolds number in the range of [0.11,0.45] & [2.69,172.96] respectively.