| description abstract | Electrohydrodynamic (EHD) printing is an additive manufacturing technique capable of microscale and nanoscale structures for biomedical, aerospace, and electronic applications. To realize stable printing at its full resolution, the monitoring of jetting behavior while printing and optimization of the printing process are necessary. Various machine vision control schemes have been developed for EHD printing. However, in-line machine vision systems are currently limited because only limited information can be captured in situ toward quality assurance and process optimization. In this article, we presented a machine learning-embedded machine vision control scheme that is able to characterize jetting and recognize the printing quality by using only low-resolution observations of the Taylor Cone. An innovative approach was introduced to identify and measure cone-jet behavior using low-fidelity image data at various applied voltage levels, stand-off distances, and printing speeds. The scaling law between voltages and the line widths enables quality prediction of final printed patterns. A voting ensemble composed of k-nearest neighbor (KNN), classification and regression tree (CART), random forest, logistic regression, gradient boost classifier, and bagging models was employed with optimized hyperparameters to classify the jets to their corresponding applied voltages, achieving an 88.43% accuracy on new experimental data. These findings demonstrate that it is possible to analyze jetting status and predict high-resolution pattern dimensions by using low-fidelity data. The voltage analysis based on the in situ data will provide additional insights for system stability, and it can be used to establish the error functions for future advanced control schemes. | |
