Composite Bi2Sr2CaCu2O8-x (Bi-2212) wire has great potential as a material for high temperature superconducting magnets, due to its ability to conduct electricity without resistance. However, during heat treatment, individual Bi-2212 filaments may agglomerate or bridge, compromising wire performance. Traditional image analysis methods struggle to quantify this agglomeration because of the wide variability in filament bridging behavior—from light to fully conjoined. In this project, we apply and compare two semantic segmentation models, U-Net and SACNet, for their ability to segment and classify filaments in transverse cross-sectional images of Bi-2212 wires. Our preliminary results show that our overall pixel accuracy is about 95% while the individual filament accuracy is about 73%. The SACNet has also been adapted to operate on the UWEC BOSE supercomputing cluster, which allows higher throughput testing at a rate approximately 19 times faster than on a standard computer operating system. The process of training the model is simple and only requires editing hyperparameters within a text document. The hyperparameters are currently being assessed for their impact on the overall accuracy of the model. We hope to turn the Python-based code into a standalone software product that can be easily used by researchers without a coding background. This should allow the software to be used widely and further our understanding of the role of bridging in the performance of the wire.
