Data-driven modeling for plasma state tracking and analysis: methods and control room integration

The development of automated analysis tools is crucial to accelerate research activities in tokamak experiments [1,2]. In this setting, some relations or patterns are challenging to model explicitly from a physics basis; data-driven tools complement physics-based approaches by exploiting large datasets of measurements. We present efforts towards developing these data-driven methods and their integration into control room operation.

Specifically, we introduce methods that directly approximate a manual analysis pipeline (supervised learning) and those that identify interesting patterns to accelerate manual analysis efforts (unsupervised learning). For the former, we consider the automatic labeling of the plasma confinement state at TCV [3], and address practical needs necessary for wide applicability: uncertainty quantification and robustness to signal issues. For the latter, we aim to find a reduced representation of the operational space of TCV and its relation to disruptive patterns [4]. Finally, we present an outlook on ongoing model development efforts and the integration of these tools in the TCV control room.

[1] F. Imbeaux et al 2015 Nucl. Fusion 55 123006

[2] A. Pau et al 2023 29th IAEA FEC 20.500.14299/250373

[3] Y. Poels, C. Venturini et al 2025 Nucl. Fusion 10.1088/1741-4326/adf349

[4] Y. Poels et al 2025 Nucl. Fusion 10.1088/1741-4326/adf121