An important part of the European activities in the area of controlled thermonuclear fusion is focused on design and R&D toward a demonstration fusion power plant DEMO. With the transition of the pre-conceptual design into the conceptual design, started in 2021, questions regarding design uncertainties and their impact on machine operation have become of increased interest. Sensor fusion allows uncertainty propagation studies and combination of data from multiple sensors or diagnostics. Using Bayesian probabilistic inference, information about the joint probability distribution of a set of plasma quantities of interest is derived, based on multiple sources of data. This way, the propagation can be studied of uncertainties through complicated, nonlinear models that describe the measurement process (forward models). At the same time, synergies in the measurements from multiple diagnostics can be exploited, since each diagnostic contributes with complementary information about the quantities of interest.

This research project involves the development of sensor fusion for DEMO diagnostics. In addition, a very novel aspect of the project concerns the application of Bayesian methods for quantifying uncertainties in the response of the plasma to actuator signals set by the control system. Another key idea behind our research relates to the relatively new development of predictive maintenance for fusion applications. This involves automated, real-time analysis of measurements from sensors and diagnostics using statistical and machine learning techniques, in order to detect and predict anomalous behavior of the plasma or machine components.

In 2024, the objectives are as follows:

• Provide forward models for density diagnostics using expertise already developed in interferometry, polarimetry, and reflectometry in the DEMO team, and perform sensor fusion for the reconstruction of 1D or 2D density profiles.
• Develop a comprehensive model for 2-D plasma current and equilibrium reconstruction, considering both magnetic and kinetic diagnostics, as well as the unknown parameters. A fast reconstruction technique is foreseen in 2025-2026.
• Continue the development of techniques for fast reconstruction of Bayesian posterior distributions that can be used in a real-time setting for plasma control purposes.
• Characterization of damage to beryllium tiles under steady-state heat load stress. Develop a model to estimate the remaining useful life (RUL) of the tiles by applying image processing techniques to data acquired from an infrared camera during tests.