Machine learning based spray process quantification

Abstract

Droplet size distribution and liquid loading are two of the most important properties to characterise spray processes. For example, nitrogen oxide (NOx) formation of liquid fuel-operated combustion systems is dominated by mixture homogenisation that is heavily dependent on atomization. While phase Doppler interferometry (PDI) is the technique of choice to measure droplet size distribution, shadowgraphy enables the detection of liquid bodies of arbitrary shape, e.g. ligaments, and thus can be used to delineate primary atomization. For this purpose, a data analysis tool is developed that enables a quantitative description of droplet size and the primary break-up process simultaneously. In this context, machine learning-based (ML) object detection provides a generally applicable approach that is used here to characterise a pressure swirl water spray. The deep neural networks are trained using synthetic training data with physics-informed domain randomisation. Two different network architectures (namely Mask R-CNN and SparseInst), backbones (ResNet50 and ResNet101) and several different training approaches are evaluated. The inference accuracy of the ML models is validated against PDI measurements in terms of droplet size distribution. The best-performing ML model provides a robust detection method for a wide range of measurement conditions. The validated ML model is used to characterise the primary break-up to delineate the effect of aerodynamic forces on the atomization of a hollow cone pressure swirl spray in high momentum gaseous co-flow. The results show the largest and most deformed liquid bodies away from the centre line in regions of high co-flow velocity.