Prediction of interfacial surface in two-phase flows atomization simulations using a U-Net Convolutional Neural Network

The prediction of the complex behavior of atomization is still a major challenge. Recent work has shown that direct numerical simulation has made good progress in simulating the full range of temporal and spatial two-phase flow scales involved. However, because of the computational cost, it is currently out of reach in an industrial context. Large-Eddy Simulations (LES) or Large Scale Simulations (LSS hereinafter) significantly reduce the computational cost, but require modeling of the two-phase flow at the sub-grid level. Artificial intelligence is a promising tool to support such modeling. The aim of this work is to evaluate a data-driven modeling approach for the reconstruction of the unresolved liquid surface for large-scale atomization simulations. A preliminary test using a U-Net Convolutional Neural Network has been successfully conducted on an atomization academic test case, consisting of a liquid jet in cross-flow. A learning process has been set-up using two-phase Direct Numerical Simulations of the jet in cross flow. The trained CNN was then able to correlate the liquid volume fraction fields, filtered to a LSS scale, to the actual interfacial area density. This result is a promising first step towards the prediction of a spray granulometry within a Large Scale Simulation.