AI and the World's Hard Problems · Nuclear fusion & energy · arXiv · 2026-10-19
Multiple Instance Learning for Interpretable Plasmoid Detection in the FLARE Device
The paper applies multiple instance learning, a weakly supervised method, to detect plasmoids in magnetic reconnection experiments on the FLARE device at PPPL, using 1,152 highly imbalanced shots. On full-channel input, the MIL model showed significant gains over a convolutional recurrent neural network in PR AUC and F1 score. Its attention mechanism helps identify relevant frames and estimate the number of plasmoids in a shot.