Machine learning approaches for plasma state monitoring in Tokamaks

Data science methods from the fields of machine learning and artificial intelligence (ML/AI) offer several opportunities for enabling and accelerating progress toward the realization of fusion energy. The massive amount of data available from current operating tokamaks, together with the exponential growth of computing power and cloud computing technologies, have enabled new scenarios in the framework of advanced data analysis and processing. In particular, machine learning methods are used to maximize the information extracted from measurements, systematically fuse multiple data sources and infer quantities that are not directly measured or cannot be easily computed in real-time during the experiments. The work proposed in this thesis focuses on the reconstruction of the electron temperature 1D plasma profile from the TCV tokamak through a neural network model. An autoencoder structure with recurrent layers is employed, which succeeds in solving the reconstruction task on a difficult fusion dataset.