Predicting Disruptions in Confined Fusion Plasmas Using Deep Neural Networks

Nuclear fusion is one of the most promising sources of clean and sustainable energy, but it still remains impractical due to technical limitations. Today, the most closely studies fusion technology is a tokamak reactor. Tokamaks confine superheated plasma using magnetic fields in a torus-shaped reactor vessel. However, the magnetic confinement technology is flawed, and the longest continuous tokamak operation time for any tokamak is on a scale of minutes. The plasma has instabilities that can lead to “disruptions” – events which can damage the reactor’s components and disrupt energy production. During “kink-mode” instability, for example, the superheated plasma can come in direct contact with the reactor wall. Therefore, it is necessary to forecast disruptions and mitigate them in time. Machine learning methods turn out to be very effective at predicting disruptions. There has been work done by Harbeck et al. [2] who successfully developed a deep learning model called the fusion recurrent neural network that not only reliably predicts disruptions on the two large tokamaks, DIII-D and JET, from which they used the experimental data to train their model, but also provides good predictions on other reactors. This raises the question of whether the same deep learning principles can be applied to HBT-EP (High Beta Tokamak – Extended Pulse), the experimental fusion tokamak reactor at Columbia University, for disruption predictions. I spent the past year working in the plasma physics laboratory at Columbia University under the supervision of Dr. Michael Mauel, Dr. Jeffrey Levesque, and Yumou Wei. I explored the technique that uses machine learning to predict plasma disruptions in HBT-EP. HBT-EP is not as cutting-edge as the above-mentioned larger tokamaks, but it allows easy access to large data sets and costs much less per experiment. Deep learning stands for deep neural networks, which consist of more than one parametrized layers of networks that map data to their representations through back-propagation training. I wrote several Python programs to complete specific tasks, including downloading the signal data taken from HBT-EP machine runs, processing the data, building the neural networks, and passing the data into the neural networks. The objective is to predict plasma disruptions events in HBT-EP using neural networks. My preliminary model correctly predicts disruptions 58% of the time, showing the potential of neural networks to improve HBT-EP’s reliability, and offering the distant promise of a sustained fusion reaction