Modeling 2D unsteady flows at moderate Reynolds numbers using a 3D convolutional neural network and a mixture of experts

We introduce MoE-Bolt (Mixture of Experts for lattice Boltzman), a novel neural network approach for predicting the unsteady state of fluid flow past a cylinder. We modeled the problem as a sequence prediction where 8 time steps previous to time t were used to predict the velocity fields of time t. With Reynolds numbers in the training set from 138 to 196, the problem was difficult because the flow was in an unsteady-state. We used a mixture of experts (MoE) to work cooperatively on solving the problem. The advantage of this cooperation is that the computing domain was decomposed without human intervention. When 4 experts were used our solution exhibited a 15 decibel improvement in the signal to noise ratio when compared to the single expert configuration. Our results and analyses show that MoE-Bolt is an effective approach for unsteady flows and it is a stepping stone for predicting flow fields at all time instants without using data from the simulation.