Study On Mechanism Of Primary Atomization Of A Pulsed Liquid Jet And Its Vortex Identification

To investigate the breakup process of jet atomization and the influence of disturbance on the jet,the volume of fluid(VOF)method and the adaptive algorithm based on tree data structure are used to simulate the atomization process,and the model based on deep learning is used to effectively identify the vortex of the jet flow field.For high-speed,non-viscous,incompressible two-phase flow,the tip,the liquid filaments and the droplets in undisturbed liquid jet change continuously with the evolution of the jet time.First,the tip appears in a mushroom shape,then the liquid filaments appear,and gradually become a thin-net shape.The liquid filaments slowly become small droplets,and the droplets spread upstream around the tip.The perturbation frequency forces on the jet produces a periodic wave on the surface of the liquid column.Compared with the steady velocity jet,the liquid filaments form droplets earlier,the tip is flatter,and the number of droplets is larger.Slowly increase the vibration frequency from small to large near the fundamental frequency,the jet broken length will decrease first and then increase.The non-perturbed length(L),probability density function(PDF),Sauter mean diameter(SMD)behavior different in the low-middle frequency and high frequency phase.Small-scale amplitude disturbances have little effect on the non-perturbed length(L).The formation,development and instability of liquid jet surface waves under the action of forced disturbance have a great relationship with vortex kinematics.The identification of vortex core structure and the analysis of its evolution characteristics are of great significance for the study of jet flow field.The paper proposes a deep learning network structure called Vortex Net,which has the ability to quickly identify vortex core regions from snapshots of jet flow field velocity clouds at any time,which is suitable for the identification of vortex structures of different scales.Combining deep learning models with Dynamic Mode Decomposition can improve the model's recognition rate.Extracting the main mode of DMD with the continuous high-sampling-rate velocity cloud image snapshots and comparing with the vortex core area,it is found that the vortex core accompanies the extreme area of the main mode,and the area far from the extreme value of the main mode does not exist eddy.Adding this constraint to the loss function of the deep learning model for parameter training significantly improves the model recognition accuracy and reduces the model error rate.