Neural network-based analysis of charge exchange spectra in KSTAR

Charge exchange spectroscopy (CES) is widely used to measure ion temperature, density, and rotation velocity profiles in various fusion devices. Conventional CES diagnostics typically rely on beam modulation techniques to separate active charge exchange components from raw spectral signals. However, in KSTAR, beam modulation is performed using the main heating beam rather than a dedicated diagnostic beam, which may potentially cause plasma disturbances. To circumvent the need for this perturbative technique during analysis, we have developed a neural network-based model (NN-CES) that infers ion temperature and toroidal rotation velocity directly from raw measured spectra. The model employs a physics-constrained architecture that optimizes prediction through combined parameter and spectral shape consistency losses. It is trained on an extensive dataset collected under representative plasma conditions in KSTAR, including low-confinement mode (L-mode), high-confinement mode (H-mode), locked mode, and resonant magnetic perturbation (RMP) application. In this work, we demonstrate that NN-CES predictions are in good agreement with results from conventional beam modulation analysis, enabling modulation-free diagnostics with the potential for real-time plasma control. • Developed a neural network model for modulation-free charge exchange spectroscopy. • Physics-informed loss function ensures consistency with target spectral shapes. • Trained on a large, diverse KSTAR dataset ( ≈ 150,000 frames) for robustness. • Potential for real-time plasma control with continuous, modulation-free analysis.