Optimizing Tikhonov Regularization in Fast-Ion Velocity-Space Tomography Through Neural Networks

In magnetic fusion research, accurately determining fast-ion velocity distributions from measured data is vital for advancing the current understanding of fast-ion behavior. Determining the fast-ion velocity distribution involves solving an ill-posed inverse problem, typically addressed through Tikhonov regularization[1], which requires careful selection of a regularization parameter. This parameter critically influences the solution. Specialized neural networks, each designed for specific fast-ion diagnostics such as fast-ion loss detectors, imaging neutral particle analyzers, and spectroscopic measurements, accurately predict the optimal regularization parameter in the mean-square sense. These networks facilitate the rapid determination of the regularization parameter for all measurements in a discharge within seconds, streamlining the computation of Tikhonov-regularized solutions and improving the efficiency of fast-ion diagnostics data analysis.

[1] Salewski M et al 2016 Nucl. Fusion 56 106024