This Thesis addresses the use of artificial intelligence methods for the protection of nuclear fusion devices with reference to the Joint European Torus (JET) Tokamak and the Wendenstein 7-X (W7-X) Stellarator. JET is currently the world's largest operational Tokamak and the only one operated with the Deuterium-Tritium fuel, while W7-X is the world's largest and most advanced Stellarator. For the work on JET, research focused on the prediction of “disruptions”, and sudden terminations of plasma confinement. For the development and testing of machine learning classifiers, a total of 198 disrupted discharges and 219 regularly terminated discharges from JET. Convolutional Neural Networks (CNNs) were proposed to extract the spatiotemporal characteristics from plasma temperature, density and radiation profiles. Since the CNN is a supervised algorithm, it is necessary to explicitly assign a label to the time windows of the dataset during training. All segments belonging to regularly terminated discharges were labelled as 'stable'. For each disrupted discharge, the labelling of 'unstable' was performed by automatically identifying the pre-disruption phase using an algorithm developed during the PhD. The CNN performance has been evaluated using disrupted and regularly terminated discharges from a decade of JET experimental campaigns, from 2011 to 2020, showing the robustness of the algorithm. Concerning W7-X, the research involved the real-time measurement of heat fluxes on plasma-facing components. THEODOR is a code currently used at W7-X for computing heat fluxes offline. However, for heat load control, fast heat flux estimation in real-time is required. Part of the PhD work was dedicated to refactoring and optimizing the THEODOR code, with the aim of speeding up calculation times and making it compatible with real-time use. In addition, a Physics Informed Neural Network (PINN) model was proposed to bring thermal flow computation to GPUs for real-time implementation.
Machine Learning and Deep Learning applications for the protection of nuclear fusion devices
AYMERICH, ENRICO
2023-04-20
Abstract
This Thesis addresses the use of artificial intelligence methods for the protection of nuclear fusion devices with reference to the Joint European Torus (JET) Tokamak and the Wendenstein 7-X (W7-X) Stellarator. JET is currently the world's largest operational Tokamak and the only one operated with the Deuterium-Tritium fuel, while W7-X is the world's largest and most advanced Stellarator. For the work on JET, research focused on the prediction of “disruptions”, and sudden terminations of plasma confinement. For the development and testing of machine learning classifiers, a total of 198 disrupted discharges and 219 regularly terminated discharges from JET. Convolutional Neural Networks (CNNs) were proposed to extract the spatiotemporal characteristics from plasma temperature, density and radiation profiles. Since the CNN is a supervised algorithm, it is necessary to explicitly assign a label to the time windows of the dataset during training. All segments belonging to regularly terminated discharges were labelled as 'stable'. For each disrupted discharge, the labelling of 'unstable' was performed by automatically identifying the pre-disruption phase using an algorithm developed during the PhD. The CNN performance has been evaluated using disrupted and regularly terminated discharges from a decade of JET experimental campaigns, from 2011 to 2020, showing the robustness of the algorithm. Concerning W7-X, the research involved the real-time measurement of heat fluxes on plasma-facing components. THEODOR is a code currently used at W7-X for computing heat fluxes offline. However, for heat load control, fast heat flux estimation in real-time is required. Part of the PhD work was dedicated to refactoring and optimizing the THEODOR code, with the aim of speeding up calculation times and making it compatible with real-time use. In addition, a Physics Informed Neural Network (PINN) model was proposed to bring thermal flow computation to GPUs for real-time implementation.File | Dimensione | Formato | |
---|---|---|---|
Thesis_def.pdf
accesso aperto
Descrizione: Machine Learning and Deep learning applications for the protection of nuclear fusion devices
Tipologia:
Tesi di dottorato
Dimensione
9.38 MB
Formato
Adobe PDF
|
9.38 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.