Seeing in the dark - Neural networks for passive CER spectroscopy analysis

Accurate ion-temperature and plasma-rotation profiles are important for transport studies and kinetic equilibrium reconstructions in tokamaks. Conventional CER spectroscopy relies on neutral beam injection (NBI) to the plasma, and on Gaussian peak fitting to extract physically meaningful plasma properties from spectra. Without NBI, the spectrometers capture information that can originate from the core and edge of the plasma. Localized measurements are not possible anymore. As a consequence, it is not possible to compute ion temperature and rotation profiles in this case. We present a deep-learning framework that infers ion temperature and toroidal plasma rotation speed directly from raw, multi-chord CER spectra. We train a convolutional encoder-decoder model on more than 5TB of spectrograms from all DIII-D discharges, covering the full wavelength range. We use a masked loss for the training process and stochastic time-frequency masking as data augmentation. This enables the AI model to learn translation-invariant spectral features while ignoring missing labels. On a held-out test set, the model achieves an RMSE of 120 eV for ion temperature and 15 km/s for plasma rotation speed, with an R2 score of 0.95. In beam-free plasmas, the predicted rotation speed correlates with independently measured tearing-mode frequencies, confirming physical fidelity. The approach extends CER profile coverage to regimes previously inaccessible, offers a path toward multi-diagnostic spectral fusion, and is readily transferable to other devices such as KSTAR and ITER.