Shape Recognition of Fuel Atomization in Crossflow using Deep Learning

To design aerospace engines efficiently, it is important to develop numerical analysis models of atomization. However, crossflow atomization, which is applied in engines that use a premixing/pre-vaporizing lean combustion in main burner, creates very complicated fluid structures and a general model for its analysis has yet to be developed. To analyze this phenomenon, it is important to consider the difference in the required resolutions between the gas field and atomized droplets, and Euler-Lagrange coupling analysis has been proposed to address this. However, in previous studies, only sufficiently spherical droplets could be replaced by Lagrange particles, and this reduces accuracy. We therefore propose a secondary atomization model that considers droplet shape and increases the number of droplets which can be replaced. To reduce the computational cost of analyzing three-dimensional droplet shape data, we applied deep learning to droplet shape recognition and breakup behavior prediction. As a result of this shape consideration, the accuracy of predicting whether or not a droplet exhibits “breakup” increased from 65.60 % to 77.68 % compared to a prediction using the theory of previous substantial model that treats all droplets as spheres. In addition, deep learning reduced the droplet shape recognition time from 4.90 s to 0.21 s compared to a method based on geometric parameters.