Description
The precise characterization of capillary jet breakup and droplet formation is fundamental to numerous applications, from inkjet printing to biochemical assays. While high-speed imaging provides detailed temporal data on these fast-evolving phenomena, manual analysis remains a significant bottleneck, particularly for large datasets involving complex fluid behaviors. This study presents an automated, high-throughput computational framework designed to extract quantitative physical metrics from high-speed images of microfluidic jets.
Our methodology integrates a custom-trained U-Net convolutional neural network (CNN) for semantic segmentation with advanced post-processing algorithms. The pipeline addresses critical challenges in microfluidic imaging, such as motion blur during the inertial-capillary regime and the segmentation of fine viscoelastic filaments. We implement a sliding-window inference strategy to process high-aspect-ratio images without resolution loss, coupled with morphological reconstruction techniques to resolve fragmentation artifacts common in low-contrast transition zones.
To quantify the breakup dynamics, the system automatically tracks key geometric parameters, including jet length, filament width, droplet aspect ratio, and pinch-off time. A Region of Interest (ROI) filtering mechanism, combined with dimensional constraints, ensures robust tracking of the primary jet and satellite droplets, effectively mitigating noise and coalescence artifacts.
Preliminary results demonstrate the algorithm's capability to accurately distinguish between the "snap-off" breakup mode of Newtonian fluids (water, glycerol) and the "beads-on-a-string" structure characteristic of viscoelastic fluids (PEO solutions). This approach significantly reduces data processing time and provides a scalable solution for investigating the rheological properties of complex fluids under extensional flow.