Diagnostics play a pivotal role in nuclear fusion experimental reactors, supporting physical studies, modelling, and plasma control. However, most diagnostics provide limited and partial information about the plasma's status. For instance, magnetic probes measure only external magnetic fields, while interferometers, polarimeters, and bolometers deliver line-integrated measurements, necessitating specific inversion algorithms to extract local information. In the case of bolometers, tomographic inversions are particularly complex due to the variety of radiative patterns observed, with regularization equations often only weakly approximating the intricate physics involved. To address these challenges, it is essential to develop innovative algorithms that enhance the accuracy of the inversion processes, thereby ensuring reliable results for physics understanding, modelling, and plasma control. This work introduces new methodologies based on Physics-Informed Neural Networks (PINNs) to perform time-resolved emission tomography from bolometer data. These methodologies are first evaluated using synthetic cases (phantoms) and compared with one of the most advanced tomographic inversion techniques in the literature. Subsequently, they are applied to reconstruct specific radiative anomalies, such as Edge Localized Modes, Multifaceted Asymmetric Radiation from the Edge, and excessive core radiation leading to temperature hollowness at the Joint European Torus. The study demonstrates that PINNs not only enhance the overall accuracy of tomographic inversions but also offer advanced capabilities like super-resolution, data projection, and self-modelling. These features make time-resolved PINNs a valuable tool for analysing radiative patterns in transient phenomena. Although this work only considers tomography, the technology is perfectly suited to tackle any kind of inverse problem and can therefore provide significant benefits for both research and practical applications in nuclear fusion.
This paper introduces new methods using Physics-Informed Neural Networks (PINNs) to accurately reconstruct radiation patterns in tokamak fusion reactors from limited sensor data. The techniques can enhance tomographic inversions, detect transient events, and provide advanced capabilities like super-resolution and self-modelling.