A physics-informed neural network approach for surrogate modeling for PWRs LOOP events

Accurate and rapid predictions of reactor transients such as Loss of Offsite Power (LOOP) events in Pressurized Water Reactors (PWRs) is critical for reactor safety assessment. Traditional highly benchmarked codes are computationally expensive, prohibiting their direct use in fast safety assessments. To address these challenges, this study investigates the potential of physics learning based surrogate models. We developed and compared data-driven and physics-informed surrogate models for predicting transient thermal-hydraulics in a triple-loop PWR under the LOOP conditions, focusing on the evolution of primary system pressure, fuel temperature and coolant temperature. Four data-driven Deep Neural Network (DNN) architectures and three Physics-Informed Neural Networks (PINNs) variants were trained using data sets covering 36 transient scenarios. Each scenario was varied with initial reactor condition, system configuration and boundary constraints. The results show that the improved Multi-Layer Perceptron (best data driven model) achieved an R 2 of 0.7015 for primary coolant temperature ( T c ), 0.6701 for fuel temperature ( T f ) and 0.8364 for primary system pressure ( P p ) on held-out test scenarios. The Shared-encoder Multi-Head PINN improved R 2 to 0.7449 for T c , 0.6987 for T f and 0.8519 for P p . The PINN also demonstrated improved robustness to sensor measurement uncertainty, with an MAPE increase of 0.1274% under 3% Gaussian input noise, compared to 0.2954% for the DNN. The study also shows incorporating governing heat and pressure equations within the learning process improves the predictive performance on held-out scenario trajectories within the training envelope, which indicates that while PINNs offer a promising path to physical consistency, the fidelity of the underlying governing equations is paramount to achieve physically consistent predictions. This finding highlights the need for a balance between data-driven learning and physical constraints.