A dimensionality reduction algorithm for mapping tokamak operational regimes using a variational autoencoder (VAE) neural network

Nov 23 2021

A new paper by members of the Plasma Physics Lab was published in Nuclear Fusion describing the use of machine learning methods to reduce the complex multi-dimensional operation space of a tokamak fusion reactor to a much simpler representation.

In tokamak experiments, a disruption is a dramatic event which rapidly terminates a plasma discharge. During a disruption, confinement of the hot plasma inside the reactor chamber is lost, dumping its energy onto the vessel wall and potentially causing damage to the machine through the resulting magnetic force and thermal load. For this reason, the ability to reliably predict and subsequently avoid or mitigate disruptions is important for the operation of the ITER tokamak and future fusion reactors.

At the Columbia Plasma Physics Lab, PhD student Yumou Wei developed an algorithm to predict disruptions using a deep learning algorithm called a variational autoencoder (VAE). A VAE is an unsupervised learning algorithm which is capable of learning meaningful data representations in a reduced dimensional latent space. It uses the encoder-decoder neural network framework while performing stochastic sampling and other techniques in order to learn a meaningful representation. VAEs have been successfully demonstrated in numerous scientific and engineering fields such as image processing and anomaly detection as well as in the physical sciences.

To implement the VAE model, Yumou and the research group (led by Profs. Gerald Navratil and Michael Mauelcollected and compiled a dataset consisting of diagnostic signals from over 3,000 plasma discharges from the High Beta Tokamak-Extended Pulse (HBT-EP) device. This dataset was used to train a VAE model, and using the trained encoder network, the data were mapped onto a two-dimensional latent space from the original seven-dimensional signal space. Within this latent space, each plasma discharge forms a continuous trajectory, and multiple operational boundaries can be identified, corresponding to different causes of disruption. Knowledge of the plasma’s current location in relation to operational boundaries within the latent space provides an intuitive way for the machine operator to not only predict an imminent disruption but to also perform disruption avoidance by steering the trajectory away from the boundary. The group demonstrated this control technique through multiple pre-programmed control experiments performed on HBT-EP and showed this technique allowed the plasma to consistently avoid the oncoming disruption and to extend its lifetime.

A dimensionality reduction algorithm for mapping tokamak operational regimes using a variational autoencoder (VAE) neural network, Y. Wei et al 2021 Nucl. Fusion 61 126063

Images: (Above) Overview of the disruption prediction and avoidance routine implemented on HBT-EP using the VAE algorithm. Data streams coming from the data acquisition system are projected to the latent dimension using the trained encoder network to form a trajectory in the latent space. An alarm is triggered once the trajectory crosses an operational boundary, which then activates the relevant control actuator on the tokamak. (Below) PhD student, Yumou Wei

Yumou Wei stands in front of the HBT-EP experiment

Yumou Wei