Measuring the electron temperature and identifying plasma detachment using machine learning and spectroscopy.

Area of Science:

• Plasma Physics
• Machine Learning
• Spectroscopy

Background:

• Accurate measurement of electron temperature (T_e) is crucial for understanding and controlling tokamak plasma.
• Traditional diagnostics like Thomson scattering can be complex and time-consuming.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for direct electron temperature measurement from plasma emission spectra.
• To assess the ML model's capability in identifying plasma detachment in tokamak divertors.

Main Methods:

• A neural network (NN) was trained using extreme ultraviolet/vacuum ultraviolet emission spectroscopy data from the DIII-D tokamak.
• The NN was trained on 1865 time slices, validated against Thomson scattering measurements.
• A NN classifier was developed to detect detached plasma states.

Main Results:

• The NN accurately predicted electron temperature, especially below 10 eV, with a mean average error < 1 eV.
• The NN classifier achieved 99% accuracy (F₁ score of 0.96) in identifying detached states, operating 10x faster than Thomson scattering.
• Collisional radiative modeling was used to analyze model performance and predict outcomes for low-cost spectrometers.

Conclusions:

• Machine learning models, combined with spectroscopy, offer a fast and accurate method for electron temperature measurement in fusion devices.
• Low-cost spectrometers integrated with ML can enhance existing diagnostics and serve as independent measurement tools.
• This approach demonstrates a proof-of-principle for cost-effective, high-performance plasma diagnostics.