20 novembre 2024
- Plasma Instability Identification Through Machine Learning / Enrique Zapata Cornejo PhD Defense
Doctorant : Enrique Zapata Cornejo
Date : 20 novembre 2024 à 13h30, amphi 3, Centrale Méditerranée au 38 Rue Frédéric Joliot Curie 13451 Marseille
Abstract:
This thesis presents the development of advanced data-driven techniques to automate the detection and classification of plasma modes in fusion experiments, focusing particularly on Alfvén instabilities and magnetohydrodynamic (MHD) modes.
The first contribution of this work is the development of a sparse coding algorithm capable of identifying modes directly from raw plasma signals. This method called Elastic Random Mode Decomposition applies parallelized elastic net regression to random dictionaries of Gabor atoms, this algorithm isolates significant oscillatory components, even with noisy signals.
In addition, unsupervised learning techniques are employed to cluster MHD modes using plasma signals and mutual information, enabling the automatic classification of different oscillatory modes without needing labeled data. These feature creation steps and clustering methods offer a scalable solution for processing large datasets from fusion experiments, allowing for systematically identifying important plasma instabilities.
The thesis also explores computer vision filtering methods for feature extraction from spectrogram images. These filters are based on spectral analysis: Fourier transform, wavelet transform, and Hough transform. They improve the quality of the spectrogram data by reducing noise and undesired features, enhancing time frequency structures related to the plasma oscillations.
Furthermore, segmentation algorithms commonly used in computer vision (CV) are adapted to identify modes in spectrogram images, enabling precise segmentation of oscillatory patterns. The pipeline of CV algorithms for segmentation is the following: noise filters, ridge detector, automatic thresholding, and labeling regions.
This result might be key for systematic signal labeling, a crucial step toward automating the labeling of plasma diagnostic signals. The methods developed here provide a necessary step for future training of deep learning models, which could further enhance real-time plasma monitoring and control in fusion reactors.
Key words: MHD modes, Alfvén instabilities, machine learning, sparse, elastic net, Gabor atoms, signal analysis, unsupervised learning, computer vision, segmentation, labelling.
Jury :
VEGA Jesús CIEMAT Laboratorio Nacional de Fusion / Rapporteur
FRÉNOD Emmanuel UBS Laboratoire de Mathématiques de Bretagne Atlantique / Rapporteur
MANTSINEN Mervi Barcelona Super Computing Center Computer Applications in Science & Engineering Department / Examinatrice
REA Cristina MIT Plasma Science and Fusion Center / Examinatrice
VERDOOLAEGE Geert Gent Univ Applied Physics Department / Examinateur
GRANDGIRARD Virginie CEA IRFM / Président de Jury
ZARZOSO David CNRS M2P2 / Directeur de thèse
PINCHES Simon ITER Plasma Modelling & Analysis Section / Co-Directeur de Thése
