Piccione, Andrea; (2022) Physics-guided machine learning approaches to predict stability properties of fusion plasmas. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Piccione_PhD_thesis_final.pdf - Accepted Version Download (34MB) | Preview

Abstract

Disruption prediction and avoidance is a critical need for next-step tokamaks such as the International Thermonuclear Experimental Reactor (ITER). The Disruption Event Characterization and Forecasting Code (DECAF) is a framework used to fully determine chains of events, such as magnetohydrodynamic (MHD) instabilities, that can lead to disruptions. In this thesis, several interpretable and physics-guided machine learning techniques (ML) to forecast the onset of resistive wall modes (RWM) in spherical tokamaks have been developed and incorporated into DECAF. The new DECAF model operates in a multi-step fashion by analysing the ideal stability properties and then by including kinetic effects on RWM stability. First, a random forest regressor (RFR) and a neural network (NN) ensemble are employed to reproduce the change in plasma potential energy without wall effects, δWno-wall, computed by the DCON ideal stability code for a large database of equilibria from the National Spherical Torus Experiment (NSTX). Moreover, outputs from the ML models are reduced and manipulated to get an estimation of the no-wall β limit, βno-wall, (where β is the ratio of plasma pressure to magnetic confinement field pressure). This exercise shows that the ML models are able to improve previous DECAF characterisation of stable and unstable equilibria and achieve accuracies within 85-88%, depending on the chosen level of interpretability. The physics guidance imposed on the NN objective function allowed for transferability outside the training domain by testing the algorithm on discharges from the Mega Ampere Spherical Tokamak (MAST). The estimated βno-wall and other important plasma characteristics, such as rotation, collisionality and low frequency MHD activity, are used as input to a customised random forest (RF) classifier to predict RWM stability for a set of human-labeled NSTX discharges. The proposed approach is real-time compatible and outperforms classical cost-sensitive methods by achieving a true positive rate (TPR) up to 90%, while also resulting in a threefold reduction in the training time. Finally, a model-agnostic method based on counterfactual explanations is developed in order to further understand the model's predictions. Good agreement is found between the model's decision and the rules imposed by physics expectation. These results also motivate the usage of counterfactuals to simulate real-time control by generating the βN levels that would keep the RWM stable.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item