Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at TCV.

Area of Science:

• Fusion energy research
• Plasma physics
• Scientific Machine Learning (SciML)

Background:

• Tokamak rampdown phases are challenging to simulate and prone to plasma instabilities.
• Existing methods struggle to predict and mitigate these instabilities, risking operational disruptions.

Purpose of the Study:

• To develop a predictive model for tokamak plasma dynamics during rampdown using SciML.
• To design control strategies that avoid plasma instability limits.
• To demonstrate the model's effectiveness in real-world fusion experiments.

Main Methods:

• Developed a neural state-space model (NSSM) combining physics-based and data-driven approaches.
• Trained the NSSM on a dataset of 311 Tokamak à Configuration Variable (TCV) pulses.
• Applied reinforcement learning (RL) to optimize plasma trajectories for stability.
• Validated the NSSM through high-performance experiments at TCV.

Main Results:

• The NSSM successfully learned plasma dynamics from a modest dataset, including reactor-relevant high-performance regimes.
• RL-guided trajectories demonstrated statistically significant improvements in operational metrics.
• A predict-first experiment showed the NSSM's capability for controlled extrapolation, increasing plasma current by 20%.

Conclusions:

• The developed SciML approach enhances tokamak control by providing robustness against uncertainty.
• This method demonstrates the practical relevance of SciML for advancing fusion energy experiments.
• The NSSM paves the way for more stable and efficient tokamak operations.