Fig. 2: Overview of the methodology and key statistical results.
From: Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV
a The training data distribution, with a modest dataset size at low performance and very few shots in the relevant high-performance regime. The target scenarios for this work at 140 kA and 170 kA with high normalized performance are shown. b Depiction of the dynamics model training method, which involves comparing results from forward simulation of an NSSM against experimental data to compute the gradient of loss with respect to model parameters. c Depiction of the trajectory optimization process. In addition to the trained dynamics model, the reinforcement learning (RL) training environment is defined by a reward function specifying the desired goal and a set of random variables that training environments are parallelized against to find a trajectory that has robustness to uncertainties and off-normal events. d Scatter plot of plasma current, Ip, and stored energy, Wtot, at time of plasma termination. Bottom-right table shows p-values from the Mann-Whitney U test comparing performance of experimental shots, with and without debug shots included, relative to the control set of all shots in the database with βN > 1.5.
