It was pointed out that a small size obstacle in such level of 5 % in width results in a huge enhancement of local electric field (E) due to strong non-linearity of electric field vs. current density (E-J) characteristics [1]. Actually, such spatial homogeneity is one of the most important requirements of practical high-temperature superconducting (HTS) wires, and all the rear-earth based HTS (REBCO) coated conductors (CCs) at present are usually tested before shipment by a continuous Ic measurement using TapestarTM at 77 K in order to clarify the defect-free quality. In our previous study, however, we found the existence of remaining local obstacles within the level of Ic fluctuation. We had succeeded in detecting local obstacles from the image classification analysis of in-plane 2D critical current density (Jc) map obtained from high speed reel-to-reel scanning Hall probe microscopy and machine learning in a 200-m-long CC by Pulse Laser Deposition (PLD) method [2, 3]. This sort of local inhomogeneity could not be detected by the standard 1D Ic characterization by the TapestarTM measurements because of a fluctuated Ic with a peak-to-peak variation more than 10 % of average Ic in the superconducting layer itself. This type of obstacles may become more severely influence on the stability at low temperature region where n-index becomes large especially in REBCO CCs. It is relevant to grasp the behavior of these spatial non-uniformity. To collect more detailed information such as size, position, and statistical behavior of the defects, we have developed object detection model in this study. We recognized two types of obstacles, i.e., an isolated obstacle and a cluster of obstacles. By this object detection model, we have succeeded in identifying local obstacles with a bounding box at the obstacle. From the coordinate of the bounding boxes, we estimated the size and position of the obstacles. We found that the size of isolated obstacles distributes less than 8 mm, whereas the size of cluster type is from 4 mm to more than 40 mm. Furthermore, we also applied this methodology to another 18-m-long CC fabricated by reactive co-evaporation by deposition and reaction (RCEDR) method. We found that some of the obstacles in the RCEDR sample could not be identified by the object detection model trained by the data from PLD sample. Namely, there are process dependent images for the obstacles. We added these obstacle image appears only at the RCEDR sample as a new obstacle group and carried out transfer learning of the model. The new model allows us to detect obstacles in both PLD CC and RCEDR CC. The length of the RCEDR-type obstacles are less than 8.5 mm, which is shorter than that in PLD CCs. In summary, we combined high-resolution high-speed magnetic microscopy and machine learning for the study of local obstacles in long-length CC. Object detection model allows us to collect detailed information on the obstacles in CC fabricated by different processes.
[1] M. Friesen and A. Gurevich, Physical Review B 63, 064521 (2001)
[2] K. Higashikawa et al., SuST. 33 064005 (2020).
[3] N. Somjaijaroen et al., IEEE TAS.32 6601504 (2022)
This work was supported by the JSPS KAKENHI Grant Number JP19H05617 and JP23K13368.