Deep learning based spraying pattern recognition and prediction for electrohydrodynamic system

Abstract

Effective recognition and prediction of spraying patterns for electrohydrodynamic (EHD) process are extremely important for its applications in high quality micro/nanoparticles preparation, chip coating, droplet-reactor design, and high precision printing, etc. In this study, six distinct spray patterns, namely dripping, spindle, cone-jet, rotational jet, atomization, and skew jet-atomization, were classified through experiments. Subsequently, 30,000 images were obtained to train a convolutional neural network (CNN) model for recognizing EHD spraying patterns, which exhibited a remarkable accuracy of 99.80%. The CNN model was used to recognize the patterns across a range of experimental variables. Dimensionless groups were established and the generalized spraying pattern maps were drawn efficiently via the model. Finally, a database consisting of 11,650 experimental data points was constructed to train a deep neural network (DNN) model, aiming to reduce the number of experiments. The DNN model with an accuracy of 95.88% was employed to predict the spraying patterns, by which a rapid but comprehensive analysis of the impact of different conditions was achieved.