Abstract
There has been a significant increase in thermoplastic composite applications in aerospace in recent years. Semi-crystalline thermoplastics like polyetheretherketone (PEEK) nucleate and grow crystalline structures called spherulites. Addition of carbon fibers forms a crystal structure called transcrystallinity. Processing parameters such as thermal uniformity, dwell time, and temperature history influence crystallinity and crystal morphology in thermoplastic composites. To investigate crystal morphology, neat PEEK and PEEK-carbon fiber composite samples are melted and cooled down while observed under a polarizing light microscope with a heating stage. A classification machine learning model based on the YOLOv3 algorithm is developed and applied to microscopy images to determine spherulite size and count in real-time. A segmentation model based on the U-Net algorithm is used to separate spherulites that have grown and impinged. The impingement of these crystal structures creates boundaries that are difficult to detect in the micrograph. The U-Net model is also used to identify transcrystalline regions that nucleate from carbon fibers. The use of these models in situ provides a unique real-time perspective of the kinetics of different crystals that form in varying processing parameters, part geometries, and reinforcement types.
