EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

AI-based plasma equilibrium parameter inference in KSTAR for DEMO-relevant reactors using only ex-vessel magnetic diagnostics

Jeongyeon Nam, Jeongwon Lee, Jun-Gyo Bak, Jayhyun Kim, Hyunsun Han, Yong Un Nam

Nuclear Fusion · 01.01.2027 · 67 (1) · 016022

DOI: 10.1088/1741-4326/aeb396 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Nuclear Fusion, online erschienen am 04.11.2026. Verfasst von Jeongyeon Nam und 5 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Ministry of Science and ICT, South Korea.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study proposes a deep neural network for high-fidelity equilibrium parameter inference, serving as a surrogate of the magnetic EFIT reconstruction and relying exclusively on ex-vessel magnetic signals. The objective is to circumvent the long-term integrity constraints of in-vessel magnetic diagnostics in the nuclear environment of the Demonstration Fusion Power Reactor (DEMO)-relevant devices. Using KSTAR data (2021–2024), a baseline model trained solely on measured ex-vessel magnetic diagnostics (the Measurement-only model) is compared with an Augmented model, which addresses informational sparsity via virtual-sensor signals computed from the GEQDSK. The Augmented model achieves predictive fidelity comparable to off-line EFIT, with a coefficient of determination ( R 2 ) above 0.99 and a range-normalized root-mean-squared error (NRMSE) of 0.332 % averaged over all parameters, against 2.385 % for the Measurement-only model. Because the virtual-sensor inputs derive from the EFIT solutions that define the targets, these figures quantify the surrogate fidelity with respect to the magnetic EFIT and constitute an upper bound on the accuracy attainable with physically installed ex-vessel probes, not that of an independent reconstruction. As the reference floor for the Measurement-only model, the magnetic fitting residuals of the parent EFIT reconstructions against the measured probe signals are quantified, yielding pooled range-normalized residuals of 1.55 % – 3.87 % per campaign with structured, campaign-persistent components that random noise perturbations do not emulate. Robust accuracy is maintained during neutral beam injection (NBI)-induced transients, and Shapley Additive exPlanations (SHAP)-based attribution verifies that the Augmented model captures the underlying magnetohydrodynamic physics, specifically the Shafranov shift. With a median single forward pass inference latency of 3.47 ms (99th percentile: 3.53 ms), the framework is a viable surrogate for conventional numerical solvers. In conclusion, this study demonstrates that an idealized, EFIT-consistent realization of a densified ex-vessel sensor configuration retains sufficient information for high-precision equilibrium parameter inference, provided that magnetic probes are physically installed at the designated locations. This establishes the information-content prerequisite for ex-vessel-only inference in DEMO. The accuracy attainable from physically installed probes, subject to conductor-model, calibration and structured measurement-model errors, is not established and requires validation against measured ex-vessel signals.

Förderung

  • Ministry of Science and ICT, South Korea

Zitieren

APA

Nam, J., Lee, J., Bak, J., Kim, J., Han, H., & Nam, Y. U. (2027). AI-based plasma equilibrium parameter inference in KSTAR for DEMO-relevant reactors using only ex-vessel magnetic diagnostics. Nuclear Fusion, 67(1), 016022. https://doi.org/10.1088/1741-4326/aeb396

Vancouver

Nam J, Lee J, Bak J, Kim J, Han H, Nam YU. AI-based plasma equilibrium parameter inference in KSTAR for DEMO-relevant reactors using only ex-vessel magnetic diagnostics. Nuclear Fusion. 2027;67(1):016022. doi:10.1088/1741-4326/aeb396

BibTeX
@article{Nam2027,
  author = {Nam, Jeongyeon and Lee, Jeongwon and Bak, Jun-Gyo and Kim, Jayhyun and Han, Hyunsun and Nam, Yong Un},
  title = {AI-based plasma equilibrium parameter inference in KSTAR for DEMO-relevant reactors using only ex-vessel magnetic diagnostics},
  journal = {Nuclear Fusion},
  year = {2027},
  volume = {67},
  number = {1},
  pages = {016022},
  doi = {10.1088/1741-4326/aeb396}
}
Kostenlos starten →