Data-driven reduced order modelling with malfunctioning sensors recovery applied to the Molten Salt Reactor case

This work presents the use of two Data-Driven Reduced Order Modelling techniques in predicting the transient response of a Molten Salt Fast Reactor when one or more sensors fail and, thus, provide wrong information; Supervised Machine Learning techniques are used to compensate for the failed sensors. Data-Driven Reduced Order Modelling integrate the physical knowledge contained in high-fidelity mathematical models with that coming from data measured on the actual system. This enables refining and updating the mathematical model, and address the challenges related to local-only observations, allowing for global state estimation. These methods are of interest when both sources of information are present, albeit incomplete, as is the case of the Molten Salt Fast Reactor. In these designs, typically operating in the fast neutron spectrum, the fuel is liquid, and no solid structures are foreseen in the core, thus making sensing and monitoring of safety-critical parameters and quantities quite challenging. Additionally, most literature studies on Data-Driven Reduced Order Modelling take the experimental observations as (noisy) ground-truth: very few works consider the case in which sensor fail or malfunction, and how this affect the state estimation.