Presentation
P54 - Disruption Forecasting with Uncertainty Quantification for Fusion Plasmas with Deep Ensembles
Presenter
DescriptionWe describe the extension of our existing deep learning model for the prediction of plasma instabilities in tokamak fusion reactors (FRNN) to ensembles of models, automatically constructed and trained using the DeepHyper framework for AutoML. Black-box optimization is performed on the long short-term memory (LSTM) network's hyperparameters using a distributed Bayesian optimization method on a DGX A100 machine. The automated search of thousands of candidate models yields more than a hundred networks with significantly improved predictive accuracy relative to the original baseline FRNN. Several ensembles of these top performing networks are constructed using different techniques, including a greedy, top-k, and gradient-based selection criteria. The diversity within an ensemble provides robust uncertainty quantification of the overall prediction of plasma stability. The prediction's total uncertainty is decomposed into epistemic and aleatoric uncertainty; the practical value of such a decomposition for tokamak machine operators and plasma control systems is discussed.
TimeTuesday, June 2719:30 - 21:30 CEST
LocationHall
SessionPoster Session and Reception
Event Type
Poster