Reconstructing patterns of missing fuel pins in PWR spent nuclear fuel assemblies via CNNs

Convolutional Neural Networks (CNNs) are investigated for identifying and localizing missing fuel pins in 17 × 17 PWR spent nuclear fuel assemblies under partial defect scenarios, where selected fuel pins are replaced with dummy pins. A CNN based on the Inception architecture is developed and trained using Monte Carlo–simulated data. Thermal neutron flux, fast neutron flux, and their spatial gradient are tallied at the 24 guide tube positions and used as input features. In return, the network provides a binary classification for each fuel pin, (0) for intact pins and (1) for replaced pins. Results show that the network performs better at identifying intact fuel pins than replaced ones. Incorporating the x and y components of the flux gradient alongside the scalar flux significantly improves the classification of both intact and replaced pins. A more detailed evaluation reveals that, despite strong global F1 scores, substantial variability remains in the detection of replaced fuel pins across different scenarios. However, a proximity-based classification criterion shows that the CNN reliably identifies the correct region of diversion for 90% of the cases. The impact of fuel burnup is also examined, with results indicating that models trained using fast neutron flux features and their gradients exhibit more stable performance and reduced sensitivity to burnup variations.