Operational space mapping on HBT-EP and DIII-D using Variational Autoencoder (VAE) neural networks

A Variational Autoencoder (VAE) is a type of unsupervised neural network which is able to learn meaningful data representations in a reduced dimension. We present an application in identifying the operational boundary of tokamak experiments. In contrast to disruption prediction by supervised learning algorithms, a VAE maps the input signals onto a low-dimensional latent space by their similarities with neighboring samples, creating a smooth operational space map in which individual shots form continuous trajectories. By projecting the operational parameters onto the same space, this provides an intuitive way for the operator to perform disruption avoidance using a relevant control actuator as a discharge approaches a stability boundary. We implemented a VAE using a dataset of over 3000 shots from HBT-EP and found it to be capable of forming a continuous operational space map and identifying the operational boundaries using a pre-specified warning time window. Pre-programmed control experiments were conducted to demonstrate the control technique using HBT-EP's saddle coils as a horizontal position actuator, showing the ability to avoid the oncoming disruptive event and extend the plasma's duration. The same analysis is presented using a selection of DIII-D signals and discharges.