Enabling Robust Trajectory and Control Design for Tokamaks with Scientific Machine Learning
Name
wang-allenw-phd-aeroastro-2026-thesis.pdf
Size
22.98 MB
Format
Adobe PDF
Checksum (MD5)
f600855136d76f32c1f8e625b8869998
Author(s)
Wang, Allen M.
Advisor(s)
Rea, Cristina
Fan, Chuchu
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Upcoming breakeven tokamaks like SPARC and ITER will require highly robust scenario, trajectory, and control design to achieve their full performance while avoiding operational delays due to heat flux and plasma disruptions. In preparation for SPARC operations, this thesis develops new control and data-driven simulation techniques with experimental demonstrations at the TCV tokamak. The work is organized around three main thrusts: 1) leveraging scientific machine learning (SciML) techniques to learn plasma kinetic dynamics from a modest experimental dataset, 2) applying reinforcement learning (RL) to the learned dynamics model to design robust non-disruptive ramp-down trajectories with a priori predictions of plasma dynamics for a scenario slightly out-of-distribution, and 3) developing a real-time adaptive magnetic controller centered around an inverse Grad-Shafranov neural network. A key strategy for managing the risks associated with plasma disruptions is to perform a ramp-down of the plasma current and stored energy. However, doing so often pushes the plasma closer to multiple instability limits. This problem is particularly acute for reactorrelevant high-performance plasmas, which can operate close to multiple stability limits to maximize fusion gain. Thus, robust trajectory design is desirable, but current simulation capabilities are limited in their predictiveness of this highly transient phase and typically do not have the speed and parallelism necessary to design for robustness against uncertainties. Thus, this thesis leverages advances in SciML to combine physics with data-driven models, developing a neural state-space model (NSSM) that predicts plasma dynamics during Tokamak à Configuration Variable (TCV) ramp-downs. The NSSM efficiently learns dynamics from a modest dataset of 311 pulses with only five pulses in a reactor-relevant high-performance regime. The NSSM is parallelized across uncertainties, and reinforcement learning (RL) is applied to design trajectories that avoid instability limits. High-performance experiments at TCV show statistically significant improvements in relevant metrics. A predict-first experiment, increasing plasma current by 20% from baseline, demonstrates the NSSM’s ability to make small extrapolations. In addition to feed-forward trajectory design, it may be desirable to have the plasma control system (PCS) adapt the plasma current and shape trajectory in real-time in response to real-time conditions to manage heat flux and off-normal events (ONEs). In addition, the plasma current and shape must be controlled to a high level of precision at all times to avoid excessive heating of plasma facing components (PFCs). Aiming to address both issues, a new approach to tokamak shape control is developed and demonstrated in experiment at TCV. The key insight is that a real-time capable inverse Grad-Shafranov solver would be able to adjust coil current setpoints to minimize control error, and even make significant shape changes in response to requests from other controllers. To satisfy real-time requirements, a neural network is trained on an offline physics-based solver and used in real-time. The neural network integrates naturally with existing classical control solutions by providing setpoints to lower level controllers. Experiments at TCV demonstrate precise control, with minimal disturbance due to auxiliary heating, despite the absence of explicit shape feedback. A partial demonstration of adaptive strike-point sweeping and response to an off-normal alarm demonstrates new advanced control capabilities this controller enables. A sufficiently fast and numerically robust physics-based solver can eventually augment or replace the neural network, providing a clear path to ameliorate concerns with using a neural network for real-time control. The experimental results, which primarily use profile changes to correct shape errors, also indicate a potential path towards accurate shape control with reduced diagnostics.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
