Machine learning approach for optimizing drop formation in inkjet printing

Abstract

Inkjet printing has involved a challenge to determine whether the ink can be printed in advance. In many pieces of research, the printability is shown from the various dimensionless numbers. However, the numbers are not enough to predict the exact drop formation, and some numbers such as Reynolds number need the drop velocity that is the unknown value before actually jetting. Furthermore, the drop formation results from the complicated interrelation between ink properties and waveform. We propose the machine learning approach for building a predictive model of jetting behavior and the method to find the waveform that generates a single drop. For making the predictive model, the image data of jetting the ink was obtained at various waveforms and the ink properties. The waveform is chosen at the range of 2 – 40 μs rising time and falling time, 2 – 60 μs dwell time, and 10 – 80 V voltage. The model ink is Newtonian fluid with different Z number in the range of 2 – 54. The model predicts primary drop velocity and the cases of jetting behavior which are divided into ‘non-jetting’, ‘single drop’, and ‘multiple drops’ from the parameters of waveform and Z number. We applied five machine learning methods (Support Vector Machine, k-Nearest Neighbor, Random Forests, Extreme Gradient Boosting, and Multilayer Perceptron) and tune the hyper-parameters of each learning model to avoid overfitting and find the optimal performance of the learning models. Multilayer Perceptron model is selected to the predictive model due to show higher accuracy than other models. After building the predictive model, we constructed an algorithm to find the optimal waveform. The algorithm repeats that the waveform and Z number input the predictive model and output values are used to compute an optimal function, which is defined in this study and has a minimum value when finding optimal waveform, and modify waveform and Z number until the optimal function becomes minimum. We applied the ink that is not used in this experiment to the proposed method in order to verify the method. As a result, the recommended waveform from the method ejected a single drop.