Neural Networks for Fast-Ion Velocity-Space Tomography Using Projected Velocities

Accurate characterization of fast-ion velocity distributions from measured fast-ion data supports the optimization of performance and stability in magnetic fusion devices. Deep neural networks, effective in reconstructing fast-ion velocity distributions from fast-ion loss detector and imaging neutral particle analyzer measurements[1], also show promise when applied to spectroscopic diagnostics such as collective Thomson scattering and fast-ion D-alpha spectroscopy. Three distinct network models, each trained on synthetic data representing spectroscopic measurements for 1, 5, and 9 lines of sight with viewing angles equidistantly spaced from 5^○ to 85^○, outperform traditional Tikhonov regularization methods. Increasing the number of lines of sight in the training data improves model robustness. Performance feature importance analysis identifies central spectral features and small viewing angles as vital for accurate reconstructions.

[1] Schmidt B et al ''Neural Networks for Tomographic Reconstruction and Uncertainty Quantification of Fast-Ion Phase-Space Distributions Using FILD and INPA Measurements'' (submitted)