Unsupervised learning about 4D features of microparticle motion.

Area of Science:

• Plasma physics
• Materials science
• Machine learning

Background:

• Material clusters form in high-temperature plasmas due to plasma-wall interactions.
• These clusters vary in size (sub-microns to mm) and move rapidly, necessitating advanced diagnostics.
• High-speed imaging and tracking are crucial for studying these dynamic phenomena.

Purpose of the Study:

• To develop and apply an unsupervised machine learning technique for analyzing high-temperature microparticle motion.
• To utilize deconvolutional neural networks for feature recognition and prediction of 4D microparticle trajectories.
• To assess the efficacy of machine learning in processing large image datasets from plasma experiments.

Main Methods:

• Development of an unsupervised machine learning algorithm based on deconvolutional neural networks.
• Analysis of two-camera videos capturing microparticles from exploding wires.
• Utilizing a locally competitive algorithm to optimize image analysis dictionaries and identify feature kernels.

Main Results:

• Identified dictionary kernels as features equivalent to local velocity vectors.
• Demonstrated strong correlation between dictionary elements from different camera views.
• Confirmed satisfaction of projection geometrical constraints for identified features.

Conclusions:

• Unsupervised machine learning offers a promising approach for analyzing large image datasets in high-temperature plasma research.
• Machine learning aids in understanding plasma-wall interactions by effectively handling complex experimental data.
• The developed technique facilitates feature recognition and prediction of microparticle motion, enhancing diagnostic capabilities.