Fig. 8: A priori predictions and experimental results for the extrapolation test scenario.
From: Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV
a Action trajectories and a priori predictions of plasma dynamics during rampdown, compared to experimental results from TCV #82877 and #82878. The RL training environment accounts for uncertainty in actuation with distributions on the action trajectory; the average of the distribution (in solid black) is used for shot programming. Control of the plasma shape proved to be a challenge for this phase, an issue also observed in previous rampdown studies6,8. Shown action variables include the plasma current, Ip, neutral beam injected power, PNBI, minor radius, aminor, elongation, κ, and high-field side gap, gHFS. Shown predictions and constraints include the stored energy, Wtot, the line-averaged electron density, \({\bar{n}}_{e,20}\), poloidal beta, βp, Greenwald fraction, fGW, rotational transform at the 95% flux surface, ι95, and vertical instability growth rate, γvgr. b Post-hoc predictions of electron temperature, Te, and density, ne, profiles compared to Thomson Scattering measurements.
