Description
Reliable tokamak operation requires improved capability to monitor and control the plasma state. Kinetic profiles provide key information about the plasma and serve as essential inputs for advanced control schemes. However, practical limitations such as restricted diagnostic availability, measurement failures, and limited sampling rates and/or radial resolution are obstacles for both real-time control applications and offline analysis. These constraints motivate the development of surrogate diagnostic observers capable of complementing or replacing missing measurements, as well as providing a framework for rejecting outliers and assessing data quality.
In this work, we propose a data-driven approach for this task and demonstrate its performance on the TCV tokamak through the reconstruction of toroidal rotation and ion-temperature profiles derived from charge-exchange spectroscopy measurements. The proposed model is designed to robustly reproduce these profiles while providing an estimate of prediction uncertainty, enabling a more reliable interpretation of reconstructed signals. We deploy the model in real-time on TCV, demonstrating a first use case of the approach for future plasma control applications. Beyond its application on TCV, we extend the approach to AUG in an offline setting targeting toroidal rotation and ion-temperature profiles, with the additional goal of comparing the model performance on the two machines.