Automated UFO detection via infrared diagnostics in fusion reactors: Application to the WEST tokamak

Abstract

We present UFOund, a deep-learning-based system for the automated detection and localization of moving particles (nicknamed UFOs) in infrared thermography data from the WEST tokamak. UFOs — small particles eroded from plasma-facing components (PFCs) — pose a significant disruption risk during experimental campaigns, accounting for approximately 35% of disruptions in WEST’s March–April 2023 experiments. Our approach processes sequences of infrared frames from WEST’s infrared thermography diagnostic using a spatiotemporal convolutional neural network. The model, trained on a manually annotated dataset of 295 infrared movies, achieves a balanced accuracy of 0.78 and an F1 score of 0.67 on an unseen test set with a detection threshold of 0.95, and gives very good qualitative results during operation at WEST. We further demonstrate a neural activation-based method to extract segmentation masks and approximate particle trajectories without additional manual annotations. Since November 2024, UFOund has been integrated into WEST’s post-pulse analysis pipeline, delivering near-real-time detection across all infrared views and significantly accelerating between-pulse decision making by the PFC Protection Officers.