ML-based semantic segmentation for quantitative spray atomization description

Abstract

Fuel spray atomization in gas turbine systems significantly impacts the combustion process and thereby emission formation. Considering the necessity for quantitative description of the influence of operating conditions on the spray breakup mechanisms, a machine learning (ML) based methodology is introduced to accurately segment the dispersed liquid from the continuous gaseous phase in shadowgraphy images. The segmented images subsequently facilitate a high-level statistical analysis of gas-liquid-interface contours and ultimately instability dynamics. For this purpose, multiple ML models varying in architecture (Semantic FPN and DeepLabV3+), datasets and augmentations are benchmarked to achieve the best performance. Subsequently, the best model is validated and used to obtain conditional statistics on the detected spray contours of three different spray types (jet-in-crossflow, pressure swirl spray and prefilming airblast spray). The model showcases high robustness, transferability across spray configurations and accuracy along multiple never-seen sprays thereby illustrating the superiority of deep learning methods for scientific image segmentation tasks. Moreover, the inferred high-level statistical analysis provides novel quantitative insights into the involved turbulence-spray interactions aiding the understanding of jet, sheet and film atomization under highly turbulent flow conditions.