Deep Learning Models to Reduce Stray Light in TJ-II Thomson Scattering Diagnostic.

Area of Science:

• Nuclear Fusion Energy
• Plasma Physics
• Machine Learning Applications

Background:

• Nuclear fusion offers a sustainable energy solution for global needs.
• Thermonuclear fusion devices like TJ-II are crucial for understanding fusion processes.
• Thomson scattering (TS) is a key diagnostic for measuring plasma temperature and density.

Purpose of the Study:

• To develop a deep learning method for reducing stray light noise in TS diagnostic images.
• To improve the accuracy of plasma profile measurements affected by stray light.

Main Methods:

• Utilized a Pix2Pix neural network, a type of generative adversarial network (GAN).
• Implemented an image-to-image translation approach to convert noisy images to clean ones.
• Trained the model on TS diagnostic images from the TJ-II fusion device.

Main Results:

• The Pix2Pix model successfully reduced stray light noise in TS images.
• Achieved up to 98% noise reduction, significantly outperforming previous methods (85% on validation data).
• Enabled more reliable measurements of plasma temperature and density profiles.

Conclusions:

• Deep learning, specifically Pix2Pix GANs, offers an effective solution for stray light noise in fusion diagnostics.
• This method enhances the reliability of TS diagnostic data for fusion energy research.
• Automated noise reduction avoids manual adjustments, streamlining data processing.