Publication
High resolution multi-physics core solvers predict safety parameters on a local scale with high accuracy while simulating local heterogeneities and their consequences. To achieve this, large computational resources are needed which renders the extensive use of such novel tools challenging. Auxiliary low fidelity low-cost solvers which can model the non-dominant physical phenomena in the reactor core can address this issue. Such tools can be built with Machine Learning (ML) techniques, which are increasingly applied in the nuclear field, to help simplify complex problems. The Laboratory of Reactor Physics and Thermal-Hydraulics (LRT) at PSI is developing in collaboration with the North Carolina State University (NCSU) a high-resolution multi-physics core solver for Cartesian PWR analysis with the 3D neutron transport code nTRACER and two Machine Learning (ML) models for the calculation of the coolant and fuel properties respectively. The use of ML to predict the fuel rod temperature of a PWR during normal operation with sub-pin resolution is studied in this work. The basis for the ML model development is the subchannel code CTF. The selection of quantities of interest and the construction of the database of the relevant ML model is discussed together with all necessary steps for the training. The methodology described in this work for the development of the ML model is designed to be applied, with few modifications, effectively to different reactors and using different fuel performance codes, without significant computational costs. The ML model is evaluated with CTF quarter core simulations. The temperature difference does not exceed 1.4 0C for the cladding and 14.55 0C for the fuel with the root mean square error (RMSE) reaching a maximum of 3.47 0C for the centerline temperature, indicating a good agreement between the ML model predictions and CTF.