Detection of zirconium hydrides in transmission electron micrographs using deep neural networks

Zirconium alloys are commonly employed in nuclear power applications. Under typical operating conditions, hydrogen ingress can lead to the formation of brittle Zr hydrides in the alloy. To study this behavior, transmission electron microscopy (TEM) is routinely used to image hydrides in Zr-alloys. However, the analysis of these TEM micrographs is a complex time-consuming task. Here, we employed a mask region -based convolutional neural network (Mask R-CNN) to automate an essential part of the analysis process: the identification and annotation of hydrides. In addition, although training a neural network usually requires large training datasets (in the order of thousands of images), the proposed framework was developed using a limited training dataset with the recourse of transfer learning. This work shows that the Mask R-CNN is capable of correctly and quickly labeling thermo-mechanically cycled hydrides in TEM images of pressure tube material.

Automated summary

Zirconium alloys are widely used in nuclear power applications, but hydrogen ingress can form brittle Zr hydrides in the alloy. To study this behavior, TEM is used to image hydrides in Zr-alloys, but the analysis of these micrographs is complex and time-consuming. In this study, a Mask R-CNN was employed to automate the identification and annotation of hydrides. With limited training data and transfer learning, the Mask R-CNN accurately and quickly labeled hydrides in TEM images of pressure tube material.