Outlier classification using autoencoders: Application for fluctuation driven flows in fusion plasmas.

Area of Science:

• Plasma physics
• Fusion energy research
• Materials science

Background:

• Magnetically confined plasmas exhibit fluctuation-driven flows crucial for modeling vacuum vessel component lifetimes.
• Mirror Langmuir probes (MLPs) offer high-time-resolution plasma sampling but struggle with large-amplitude fluctuations.
• Ambiguous data classification in MLPs can reach up to 40%, impacting parameter accuracy.

Purpose of the Study:

• To develop a robust method for classifying ambiguous plasma data obtained from MLPs.
• To improve the accuracy of plasma parameter and heat flux calculations by effectively removing outliers.
• To avoid subjective threshold-based outlier identification methods.

Main Methods:

• Utilized an autoencoder (AE) to learn a low-dimensional representation of valid plasma data.
• Employed standard classifiers in the AE's latent space to categorize ambiguous data samples.
• Compared heat flux calculations after removing outliers identified by the AE versus traditional threshold methods.

Main Results:

• The AE successfully learned a low-dimensional representation distinguishing valid from ambiguous data.
• Outlier removal using the AE resulted in 5%-15% lower average radial heat fluxes compared to threshold methods.
• Triple correlation contributions to radial heat flux showed up to a 40% difference when using AE-based outlier removal.

Conclusions:

• AE-based outlier identification provides a more objective and accurate method for analyzing MLP data.
• This approach significantly refines calculations of conductive, convective, and triple-correlation heat fluxes.
• The findings contribute to better understanding plasma boundary layers and predicting component lifetimes in fusion devices.