Implementation of Artificial Intelligence at PK-4 for the Analysis of Complex Plasmas

Complex plasmas serve as a distinctive experimental framework for particle-resolved investigations of strongly coupled many-particle systems under controlled laboratory and microgravity conditions. However, experiments such as the Plasmakristallexperiment-4 generate large image datasets whose analysis is complicated by noise, limited observation times, evolving experimental conditions and inherent ambiguities between ordered and disordered states. Furthermore, the necessity of human intervention during the experiments limits repeatability and is a source of error. This dissertation examines how machine learning methods can be systematically integrated into both the analysis and operation of complex plasma experiments in a physically consistent and computationally efficient manner. The core of this cumulative dissertation comprises three publications that address complementary aspects of particle-resolved data analysis. A convolutional encoder–decoder network is employed to identify string-like particle structures in two-dimensional camera images, with network architecture and training data generation guided by the underlying electrorheological physics. An unsupervised self-organizing map is applied and optimized for robust particle tracking between consecutive frames, enabling reliable trajectory reconstruction in dense and high-velocity particle flows without manual threshold tuning. To further improve data quality, an outlier detection procedure is developed. Based on the particle trajectories, turbulence in complex plasmas is formulated as a time-resolved classification problem and investigated using a long short-term memory network operating on physically motivated, invariant trajectory features. Beyond the individual publications, this work demonstrates the practical deployment of machine learning methods directly at the experiment. The developed models are adapted to embedded hardware through compact network architectures and mixed-precision inference, enabling data analysis with minimal latency implemented into the experiment. In addition, molecular dynamics simulations are analyzed using the machine learning framework, revealing systematic dependencies of turbulent particle motion on simulation parameters. Overall, this dissertation establishes a physically motivated and extensible machine learning framework that advances automated data analysis and diagnostics and supports future adaptive control strategies in complex plasma experiments.

Collections

Dissertationen/Habilitationen

Full item page

Implementation of Artificial Intelligence at PK-4 for the Analysis of Complex Plasmas

Files

Date

Authors

Advisors/Reviewers

Further Contributors

Contributing Institutions

Publisher

Journal Title

Journal ISSN

Volume Title

Publisher

License

Quotable link

Abstract

Link to publications or other datasets

Description

Notes

Original publication in

Original publication in

Anthology

Collections

URI of original publication

Forschungsdaten

Series

Citation