Modeling secondary breakup in atomization processes via machine learning

A highly sought goal of simulations of liquid jet atomization is to accurately predict the size distribution and number density of the atomized drops. This data is crucial for prediction of the surface-to-volume ratio, influencing evaporation rate, combustion efficiency and other important quantities. For realistic jets though, there is extreme separation of scales between the dynamics of the jet bulk and the breakup processes that produce the smallest drops, making it impractical to resolve all drop breakups. Thus, models must be employed for secondary breakup. However, existing breakup models are not universal and do not account for the local and instantaneous flow field and drop geometry. We introduce a machine-learning based model to predict the outcome of under-resolved drop breakups. Training the model on outcomes from the corresponding high-resolution simulations, we predict breakup outcomes from low-resolution, under-resolved input fields. Compared to results generated by low-resolution simulations, our ML-based approach achieves higher accuracy at predicting drop breakup and predicts a resulting droplet size distribution that is closer to the ground-truth distribution. Furthermore, our method learns the importance of physically relevant features such as the Weber number.