Princeton PPPL demonstrates ML-based disruption prediction at 95% accuracy

Researchers at the Princeton Plasma Physics Laboratory published a paper in Nature Energy demonstrating a convolutional neural network that predicts tokamak plasma disruptions with 95% accuracy 50 milliseconds in advance. The model was trained on a combined DIII-D and JET dataset.

Disruption prediction is critical for reactor-scale tokamaks, where uncontrolled plasma collapse can cause damaging electromagnetic loads on the first wall. Lead author William Tang said the technique is portable across machines — a property earlier ML approaches lacked.