Characterization of Droplet Shapes from Atomizing Spray Liquid Jet in Crossflow via Machine Learning

Atomization is the process of rapidly breaking down a liquid stream into droplets. Atomization is widely deployed in areas where rapid heat and mass transfer is desired, including for fuel injection and combustion in aerospace engines (termed spray combustion). To accurately predict spray combustion, the dispersion of fuel droplets and the droplet evaporation rates must be understood. However, models for the droplets are incomplete in that they consider all droplets to be perfect spheres rather than modeling the actual droplet shape. This leads to error in droplet dispersion prediction (via inaccurate drag calculations) and errors in prediction of the evaporation rate (via inaccurate prediction of heat/mass transfer). Rather than make assumptions about droplet shape, this study extracts the actual shape of droplets in sprays using machine learning. The study uses high-fidelity direct numerical simulation to investigate the atomization of the liquid stream. Droplets are identified, and their shape is characterized through a small set of geometric features. Cluster analysis (DBSCAN) is used to categorize droplet shapes and to identify the most common shapes. Given the 2D nature of the simulation, the analysis focuses on the in-plane aspect ratio. The study reveals that the spray population is dominated by quasi-spherical droplets, which constitute over 70% of the identified structures. However, a significant portion of the mass exists as highly elongated ligaments and deformed ellipses, which represent the transient breakup stages. This classification provides a quantitative basis for improving spray models by distinguishing between stable product droplets and transient ligament structures. This initial work is performed in 2D and serves as a proof of concept for future atomization studies in 3D.