Innovative reactor technologies in the framework of Generation IV are usually characterised by harsher and more hostile environments compared to standard nuclear systems, for instance, due to the liquid nature of the fuel or the adoption of liquid salt and molten as coolant. This framework poses more challenges in the monitoring of the system itself; since placing sensors inside the reactor itself is a nearly impossible task, it is crucial to study innovative methods able to infer the behaviour of the reactor from out-core sparse measurements. Recently, novel approaches have been developed able to combine in a quick, reliable and efficient way two different sources of information characterising the system, namely mathematical models and real data (i.e., measurements); these methods fall into the Data-Driven Reduced Order Modelling framework. Within this idea, Machine Learning algorithms can be easily integrated to learn the missing physics or the dynamics of the problem, in particular, they can be adapted to generate surrogate models able to map the out-core measurements of a simple field (e.g., neutron flux and temperature) to the dynamics of un-observable complex fields (precursors concentration and velocity). This work applies this idea to a Molten Salt Fast Reactor during an accidental transient, coupling the Generalised Empirical Interpolation Method with Gaussian Process Regression to indirectly reconstruct the unobservable fields.
 

Supervised learning,Reduced order modelling,model bias correction,molten salt reactor,model-data integration