Machine learning assisted characterisation and prediction of droplet distributions in a liquid jet in cross-flow

Abstract

In this paper, an artificial neural network (ANN) is trained with large eddy simulation (LES) data to predict the droplet size distribution (DSD) from the primary atomisation of a liquid jet in gaseous cross-flow (JIC), in terms of the Weber number (), momentum flux ratio () and density ratio. The JIC is simulated considering three (250, 500, 1000), (1, 5, 10), and density ratios (10, 100, 1000), respectively. The accuracy of the simulations is enhanced by including the injector geometry as well. The training data are obtained using LES with a stochastic fields transported-probability density function (PDF) method. We initially provide a physical analysis of the droplet distributions observed. We find that for lower density ratios, the resulting spray is mostly dominated by , influencing the main mechanisms governing the break-up process, which change the DSD shape. This dual mechanism is not present when increasing the density ratio. In the second part of the work, we build an ANN model (based on a multi-layered perceptron) using the DSDs from the LES as a train-and-test dataset, to predict at the end the full DSD for the JIC given as input the three non-dimensional parameters. The DSD from the trained ANN is found to be a good fit for the range investigated, predicting both the stochastic nature and change in shape of the droplet populations upon varying the input parameters. The developed model is intended to enhance future simulations of secondary atomisation in Eulerian-Lagrangian frameworks by providing the initial DSDs.