The availability of large amounts of data has made possible the latest advancements in Machine Learning and Artificial Intelligence, as they provide sufficient information for the training phase. However, the nuclear reactor field is still far from this reality for multiple reasons: the computational cost of the models for simulation, which are inherently multi-physics and thus quite complex; the scarcity of sensors and their limited lifetime, as in-vessel actuators must operate in quite harsh environmental conditions; the stringent requirements by regulators, which imply the use of detailed and complete high-fidelity models also in the design and optimisation phase, which, by its nature, falls into the multi-query scenario class. As such, the nuclear industry must still rely on physical models to provide, at least, the background information; thus, the possibility of combining the available model with data collected on the physical system is worth investigating. This investigation has a two-fold goal: improvement of the performance of the former from the computational point of view without sacrificing accuracy and performing model bias correction with the knowledge coming in real-time from the system, also for control and monitoring purposes. This work investigates this possibility by adopting the Data-Driven Reduced Order Modelling framework to perform the above; for validation purposes, the DYNASTY experimental facility built at Politecnico di Milano has been used as a test case.
 

Model-data integration,reduced order modelling,surrogate model