기계학습을 활용한 잉크젯 프린팅 웨이브폼 최적 설계

Abstract

Drop-on-demand(DOD) inkjet printing has been used in diverse areas such as printed electronics, display and tissue engineering. It forms precise drops by stimulating a short pressure pulse. The drop formation depends on relationship with inertial force, viscous force, and surface tension. The unbalanced relationship of them makes the satellite drops, which disturb stable manufacturing. Elimination satellite drops in inkjet printing is a challengeable problem because it has a complex relationship with characteristic of fluids and a waveform information. We propose a method to optimize a waveform about unknown fluid using machine learning in this thesis. The volume of drops is typically the range of 1-100 pl and the drop jetting process occurs within 1 ms. We need a system that can capture high-resolution images and record in μs unit. There are two systems to observe the jetting process. The one is cinematography method using a high-speed camera with exposure time less than 0.5 μs and frame rate exceeding 1 MHz and the other is a flash photography method due to high repeatability of fluid phenomena in the inkjet process. The flash photography is commonly used to visualize drop formation. We set up a flash photographic system composed of inkjet nozzle, camera, light source, electronics drive and digital output module, which sends pulses to them. The system can visualize drop formation and save images with an interval of 1 μs. The image analysis system which is built by ourselves extracts drop velocity and jetting behavior from sequence images. We observe drop formation and drop velocity in random waveform each fluid on this system. Z Number, which is calculated with density, viscosity and surface tension, and the parameters of the waveform are used as inputs and drop formation and drop velocity are used as outputs. We prepare range as 10 ≤ Z ≤ 54 of model fluids and select Support Vector Machine (SVM), Random Forests (RFs), k-Nearest Neighbors (kNN) and Artificial Neural Networks (ANN), which are used widely. ANN has more significantly improved performance than other machine learning models. The optimal waveform is found using feedback algorithm with prediction results. The method can find the waveform that forms single drop without ejecting ink actually. This work has the potential to analyze complex drop dynamics in inkjet printing using a machine learning method.