Emulating Operator Expertise with Neural Networks for Tokamak Stability Control

We propose a control room AI that emulates an experienced operator by using prior data to suggest adjustments that stabilize the plasma and preserve intended changes. In early-stage tokamak operations, only a limited set of scenarios is available, and small changes often cause instability or premature termination. Thus, stability often depends on the operator’s ability to interpret diagnostics and apply control adjustments [1]. Our model is trained on experimental and simulated data from the Mega Ampere Spherical Tokamak (MAST) [2] and includes two components: a neural image predictor and a control model. The predictor uses diagnostics (e.g., plasma current and density evolution [3]) from the FAIR-MAST dataset [2] to estimate plasma shape. The control model then computes the actuation needed to reach the target shape. Designed for inter-shot analysis, the model is tested on hold-out experimental data to assess predictive accuracy and generalization to realistic scenarios.

[1] J. Seo et al., Nature 626, 746 (2024).

[2] S. Jackson et al., SoftwareX, 101869 (2024).

[3] K. J. Gibson et al., PPCF 52, 124041 (2010).

[4] V. Gopakumar et al., NF 64, 056025 (2024).

[5] H. Meyer et al., Nucl. Fusion 57, 102014 (2017).