Quantification of line dislocations in FFTF irradiated HT9 cladding by deep learning method

Dislocations in crystalline materials are directedly linked to the plastic deformation and bulk mechanical properties in materials. Specifically, this work is inspired by the need for microstructural dislocation information to support a model for thermal creep of HT9 ferritic/martensitic stainless steel. Transmission electron microscopy (TEM) can directly reveal dislocations at nanoscale; however, deriving accurate dislocation statistics remains a challenge. This difficulty arises from the complex contrast mechanism and multitude of microstructural features in TEM micrographs, as well as the natural human bias in widely used manual labelling methods. Preliminary computer vision models, which employ segmentation-based neural networks and involve multiple steps of manual manipulation and analysis, often fail to detect sparse and fragmented line dislocations effectively. This work presents an end-to-end deep learning-based method for versatile dislocation line detection in TEM micrographs. By treating dislocations as edges rather than cracks or lines, we designed two modules to adapt the network from general edge detection to TEM dislocations detection. Specifically, a rectification module was developed and integrated into the basic framework to enhance the feature segmentation for continuous and non-cycle detection of TEM dislocations. Additionally, a post-processing module was embedded into the model network to filter out redundant and overlapping dislocations. The method was then used to automatically generate dislocation locations and lengths in untested TEM micrographs with high accuracy. This method has the potential to be applied to large datasets of TEM micrographs of high dislocations' density (in the order of 10¹⁴ m⁻²).