Determining 1D fast-ion velocity distribution functions from ion cyclotron emission data using deep neural networks

Area of Science:

• Plasma Physics
• Machine Learning
• Fusion Energy Research

Background:

• Ion Cyclotron Emission (ICE) is a key diagnostic for understanding fast ion behavior in fusion devices.
• Accurately inferring fast ion velocity distributions from ICE signals is crucial for plasma control and physics studies.
• Current methods for inferring velocity distributions can be computationally intensive or limited in accuracy.

Purpose of the Study:

• To develop and optimize a deep neural network model for predicting fast ion velocity distribution functions from simulated ICE signals.
• To compare the performance of the neural network against traditional methods like Tikhonov regularization.
• To assess the computational efficiency and potential applicability of the neural network to experimental data.

Main Methods:

• A two-layer deep neural network was designed to model the relationship between ICE signals and 1D velocity distribution functions.
• Network architecture and hyperparameters were fine-tuned using cross-validation and a bottom-up approach.
• Training and test datasets were generated by simulating ICE signals (s) from known velocity distributions (f) using the linear equation Wf = s.
• Performance was evaluated by comparing network predictions to true distributions and to results from 0th-order Tikhonov regularization.

Main Results:

• The deep neural network achieved higher accuracy in predicting fast ion velocity distribution functions from simulated ICE signals compared to 0th-order Tikhonov regularization.
• The neural network demonstrated significantly faster computation times.
• Qualitative differences in feature prediction accuracy were observed between the two methods.

Conclusions:

• Deep neural networks offer a promising, efficient, and accurate method for inferring fast ion velocity distributions from ICE signals.
• The developed network model can be adapted for analyzing experimental ICE data from devices like LHD.
• This AI-driven approach has the potential to enhance plasma diagnostics and control in fusion research.