Description
Scenario development operation is one of the most resource-intensive and expertise-dependent processes in magnetic confinement fusion. In current practice,
to design the desired plasma and fusion device state, researchers tune the currents of
each coil over the evolution of the discharge through an iterative trial-and-error procedure. This workflow requires significant time and limits possible optimizations. Moreover, small perturbations of discharge parameters can change the MHD equilibrium, leading to unstable or infeasible plasma evolutions.
We introduce a reinforcement learning–based framework for automated, goal-driven scenario generation. Instead of prescribing coil current waveforms, operators define high-level plasma objectives. These include target plasma shape and strike-point location, plasma current evolution ($I_p$ at flattop and ramp rate), safety factor constraints, beta optimization, ion and electron temperature profiles and minimization of poloidal flux consumption.
We employ the SAC algorithm to train a neural network controller. The policy maps the plasma state — MHD equilibrium, plasma geometric characteristics, kinetic profiles, and magnetic system state — to magnetic actuator commands, i.e., coils’ currents. In this way, low-level tuning is replaced by goal-driven optimization.
A key enabler of the approach is NSFsim---the fast Grad–Shafranov and 1D transport simulator. The simulator initializes from realistic post-breakdown limiter plasmas below 100 kA and evolves the discharge through the ramp-up phase. The reward function integrates geometric accuracy, X-point positioning, plasma current tracking, and coil flux consumption. Additionally, the training process presents randomized auxiliary heating and current drive systems. This process allows testing and optimizing the controller model, which is able to deliver realistic magnetic system commands for optimal plasma current growth, transition to the divertor configuration, and elongating the plasma under varying heating scenarios.
The framework is intrinsically device-agnostic, since the NSFsim allows for simulation
of various fusion devices, and there are only a few algorithm hyperparameters to tweak. The approach transforms scenario design from manual tuning into a semi-automated pipeline. Although expert validation remains essential, the method reduces the effort to develop scenarios and supports advanced control development and advanced experimental planning for future fusion power plants.