A Two-level Method for Modeling Fluid Movement Using a Lattice Boltzmann Model and a Convolutional Neural Network

Abstract

A new two-level method for modeling fluid movement in closed surfaces is proposed. The metod simulates an unsteady hydrodynamic process and includes two levels of description of the modeling process. The first level supports the development of the process over time and is implemented based on the Boltzmann lattice model. At the second level, for each time layer, based on the obtained velocity field, the pressure distribution is refined by modeling the solution of the Poisson equation in the working area using a convolutional neural network, which is pre-trained on a training data set formed for a given set of typical problems. A method combi¬ning both technologies is proposed, taking into account the compensation of the compressibi¬lity characteristic. The structure and features of neural network training are described. Experiments were conducted on models simulating the human digestive tract in various states. The performance of the developed method is compared with the numerical way of solving the Poisson equation.