Fundamental Insights into Additive Manufacturing with High-Speed Imaging and Machine Learning

Additive manufacturing (AM) offers great design flexibility but remains constrained by the complexity of tuning within high-dimensional, nonlinear process parameter spaces and the resource intensity of traditional simulation-based optimization. Further, AM processes involve critical transient phenomena at a 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 directly from limited experimental data, as well as reducing the need for computationally expensive physics-based models and extensive post-training experimentation. Electrohydrodynamic (EHD) jet printing is used as a representative AM process due to the interplay between processing parameters and its exhibition of complex multi-physics such as fluid dynamics, electric fields, and microfluidics. A Gaussian Process (GP) classifier is trained on 123 high-speed video sequences captured at 15,000 frames per second. It achieved 92% accuracy in identifying five printing regimes, including cone-jetting, micro-dripping, dripping, unstable, and no-print, across a 2500-6000 V and 0-5 psi parameter space. A GP regressor further predicted jet breakage velocities with an R² of 0.955, revealing consistent regime-dependent trends. A deep learning model combining a 3D convolutional encoder and transformer-based temporal predictor accurately simulated frame-by-frame droplet dynamics, achieving high performance in simulating unseen combinations of processing parameters on both qualitative and quantitative metrics. These results establish a new data-driven foundation for simulation, predictive control, and rapid or real-time optimization in complex AM processes.