Master thesis defense by Andreas Steensen Hjortsø

Abstract: This thesis investigates the applicability and limitations of physics-informed neural networks (PINNs) as surrogate models for steady-state molten salt reactor (MSR) simulation. Using the CNRS multigroup MSR benchmark simulated in Moltres as a reference model, two PINNs are developed and validated. The first, a neutronics-only model solving the multigroup diffusion k-eigenvalue problem, and the second, a partially coupled neutronics-thermal-hydraulics model with one-way temperature feedback. 

PINNs are compared to a data-driven neural network in data scarce and data rich regimes. The PINNs achieve lower relative L2 error in the low-data regime where only supervised boundary data is present. When more interior data is available the PINN is outperformed by the data-driven NN. A detailed analysis of the PINN loss landscape shows that loss terms and parameters are strongly coupled, leading to optimization challenges and difficult loss weighting. Furthermore it is determined that most of the flux error is due to amplitude, not spatial structure. Pure shape error is approximately 18 % and amplitude error ranges from 0 % to 40 % depending on neutron energy. 

The coupled model is able to predict neutron flux comparably to the neutronics-only model and pressure fields are accurate. The current coupled PINN is unable to correctly predict temperature or velocity. These results show that while PINNs are effective surrogate models for steady-state neutronics in data-scarce regimes, significant challenges remain for multiphysics coupling, emphasizing the need for further research into loss formulation and coupling strategies.