Deep Learning–Based Data-Driven Analysis of Complex Plasmas in a Direct Current Discharge
Lade...
Datum
Autor:innen
Betreuer/Gutachter
Weitere Beteiligte
Beteiligte Institutionen
Herausgeber
Zeitschriftentitel
ISSN der Zeitschrift
Bandtitel
Verlag
Zitierlink
Zusammenfassung
Complex plasmas, which are composed of electrons, ions, neutral gas atoms and micrometer sized particles, provide a unique platform for studying fundamental physical phenomena. Because the interparticle distances are comparatively large, it is possible to resolve individual particles optically. This makes it possible to investigate processes such as crystallisation, phase transitions and collective excitations. One of the experimental platforms for studying complex plasmas is the Plasma Crystal Experiment 4 (PK-4), which operates under direct current (DC) conditions. It is used on Earth and in microgravity environments, like on the International Space Station (ISS) and during parabolic flights. To reach the full potential of complex plasmas, however, robust methods must be developed to detect, track, and classify microscopic particles. Traditional image processing techniques often reach their limits in this context, especially for large datasets or under experimentally induced image noise. Therefore, modern machine learning approaches and deep neural networks offer a promising way to optimize and automate the analysis of complex plasmas, where possible.This dissertation presents a comprehensive framework for using deep learning Methods to analyze complex plasmas in the PK-4 experiment. The work is organized into three main contributions. First, a compact U-Net architecture is developed for efficient and accurate particle detection and tracking in dense plasmas. Second, WignerNet, a PointNet based model, is introduced to enable the local classification of crystalline domains using three-dimensional, Voronoi-based representations. Third, an extended graph neural network approach is used to identify more complex structures. Combining these methods greatly improves the diagnosis of complex plasmas by enabling scalable analysis at the single-particle level. In this way, the dissertation contributes both methodologically and experimentally to a deeper understanding of the dynamics and self-organization of complex plasma systems.