Machine-learning-based model for motional Stark effect calibrations

The motional Stark effect (MSE) diagnostic on the Korea Superconducting Tokamak Advanced Research (KSTAR) device requires a beam-into-gas calibration to remove the Faraday-rotation contribution to the measured polarization angle. Because the secondary-neutral emission that contaminates this calibration grows with the torus gas pressure, the calibration is obtained by measuring the polarization angle over a range of pressures at a fixed vacuum field and extrapolating to zero pressure. Such measurements, however, exist only for a finite and irregularly sampled set of the major experimental parameters-the neutral-beam energy, the toroidal magnetic field, and the beam source-so that a calibration matching a given plasma discharge frequently does not exist and the conventional remedy of a simple linear interpolation between the nearest calibrated points remains incomplete. This work develops a supervised machine-learning model that infers the deviation of the MSE-measured polarization angle from the vacuum field for an arbitrary combination of these parameters, using the beam-into-gas data accumulated over the KSTAR FY2025 and FY2026 campaigns. A two-stage scheme that first reproduces the zero-pressure extrapolation and then learns the configuration dependence of the resulting angle is found to perform best. The improvement over the conventional linear-interpolation approach is demonstrated through the magnetic-axis evolution and through the safety factor and current density profiles of off-axis current-drive discharges.