[1E07-09] Properties of Fuel and Debris, and Fuel Fabrication Technology

2:45 PM - 3:00 PM JST(5:45 AM - 6:00 AM UTC)

[1E07] A Deep-Learning Approach Towards Automated Crystalline Phase Identification in Simulated Fuel Debris

*Yifan Sun¹, Chisato Sakaguchi², Yoichi Endo², Toru Higuchi², Ken Kurosaki¹ (1. Kyoto Univ., 2. NFD)

Keywords:

X-ray diffraction,Simulated fuel debris,Phase identification,Machine learning,Convolutional neural network

Characterizing the crystalline phases of the fuel debris is a critical step towards the decommissioning the Fukushima Daiichi Nuclear Power Plant. However, conventional analysis relies on manual expert interpretation that is slow and subjective, due to the chemical complexity of these systems. To overcome this, we propose automating phase identification with deep learning, which is well suited to such high-dimensional diffraction data. Adapting the autoXRD framework to the debris-relevant U-Zr-Fe-Cr-Ni-O system, we curated a training library of 119 phase structures and 90 generated solid-solution structures. We then simulate their XRD patterns, augment them with realistic lattice strain, texture, broadening, and noise, and train a convolutional neural network for phase identification. Finally, we extend the method to multi-phase samples through an iterative subtraction workflow that sequentially identifies and removes individual phase contributions, and validate its effectiveness using measured XRD patterns of simulated fuel debris.