Investigating nozzle-based additive manufacturing dynamics with high-speed imaging and machine learning

Abstract Additive manufacturing (AM) offers great design flexibility but remains constrained by the complexity of tuning processing parameters from a vast parameter search space either by traditional trial-and-error experimental approach or expensive computational fluid dynamics simulations. Further, AM processes involve critical transient phenomena at very small timescales ranging from microsecond to millisecond, which are difficult to observe and characterize experimentally. This work presents a novel approach that combines high-speed photography with machine learning (ML) to model and predict dynamic printing behavior. We used jet printing data obtained from Electrohydrodynamic (EHD) jet printing as a representative AM process. A Gaussian Process (GP) classifier is trained on 123 high-speed video sequences captured at 15,000 frames per second accurately predicts five printing regimes, including cone-jetting, micro-dripping, dripping, unstable, and no-print, across a wide range of parameters. Further, a deep learning model combining a 3D convolutional encoder and transformer-based temporal predictor provided highly accurate quantitative estimation of frame-by-frame droplet dynamics at a computational cost orders of magnitude less than equivalent computational fluid dynamics simulations. These results establish a data-driven framework for the simulation, predictive control, and rapid optimization of EHD printing dynamics, demonstrating the potential of high-speed imaging-informed machine learning as a computationally efficient surrogate for complex AM processes.