A Deep Dive Into Disruptivity: Learning to Predict and Avoid Disruptions

In order to effectively operate high performance tokamaks, it is imperative to prevent disruptions either in real time control or in scenario planning. This poster will provide a deep dive into the disruptivity (disruptive frequency), a metric first introduced for the analysis of statistical properties of disruptions on JET [De Vries et al NF 2011]. The disruptivity is used to create maps that outline safe and unsafe regions of plasma parameter space that can be used for many applications. For operational planning, operators can look for dangerous plasma configurations or precursor events in disruptivity maps that can be avoided. In real time, control policies based on gradient descent can provide a restoring force that pushes the control system away from data-driven nonlinear boundaries. Disruptivity can be inverted to predict the time to disruption, further informing the control system on which emergency actions are viable. In simulations, disruptivity maps can create probabilities for which off-normal events can occur. An analysis of disruptivity hyperparameters, such as the method in which the disruptivity is calculated and the definition of the time of disruption, is presented. For new machines and exploration of parameter space, initial guesses for the disruptivity in the form of priors are introduced to the formalism. Analyses are conducted on Alcator C-Mod and TCV using the tokamak-agnostic disruptionStatistics (https://github.com/pkaloyannis-cfs/disruptionStatistics) code base.

Work funded by Commonwealth Fusion Systems.