Case Study: Princeton University accelerates fusion disruption prediction with NVIDIA deep learning

Speeding The Path To Fusion Energy With Deep Learning

Princeton University worked with NVIDIA to tackle one of fusion energy’s biggest challenges: predicting disruptive events in tokamak reactors quickly enough to avoid damage and downtime. The university’s researchers needed a more accurate and faster way to anticipate plasma disruptions in experiments like JET and eventually ITER, using deep learning and GPU acceleration.

Using NVIDIA Tesla P100 GPUs, CUDA, cuDNN, and the TensorFlow framework, Princeton developed the Fusion Recurrent Neural Network (FRNN) to train on large experimental datasets and forecast disruptions. The NVIDIA-powered system scaled to 200 GPUs and improved prediction performance to about 90% accuracy with less than 5% false positives, with a goal of reaching 95% true positive accuracy and 30 milliseconds or more of warning time before disruption.

Princeton University

William Tang

Princeton University’s Program in Plasma Physics and the Princeton Plasma Physics Laboratory