Enhancing multi-physics modeling in new-generation nuclear reactors using machine learning: Implementing Gaussian Process Regression for updating cross sections

Due to the complexity of the multi-physics behavior in next-generation Molten Salt Reactors (MSRs), developing high-fidelity simulation codes is essential. Despite recent advancements in developing these codes, updating cross sections as a function of state variables remains a persistent challenge. In this work, a new method for updating neutron cross sections based on Gaussian Process Regression (GPR), a supervised learning technique is developed to enhance the fidelity of the coupling between neutronic and thermal–hydraulic solvers. In this method, GPR is employed in place of the traditional logarithmic correlation (Log) to update cross sections based on fuel temperature. The GPR method is integrated into the PoliMi multi-physics code, which is developed on the OpenFOAM libraries, and is tested on two cases: the lid-driven cavity benchmark and the 2D axisymmetric MSFR. The GPR model is verified using test data generated by Monte Carlo Serpent simulations. The numerical results of this study suggest that for systems containing isotopes exhibiting significant absorption resonances (such as Th232 and U238), the method of updating cross sections has a greater impact on the predicted Keff than the choice of neutronic solver (Diffusion or SP3). Additionally, the updating method has a smaller impact on the SP3 solver compared to the Diffusion solver. The reactivity injection accident simulation results demonstrate that updating cross sections can significantly affect the predicted peak power. This is particularly important in multi-physics transient modeling for reactor safety analysis.