Surrogate Modeling Advancements in Nuclear Fuel Engineering: Bridging the Gap Between Accuracy and Efficiency

Abstract In nuclear reactor fuel engineering, the phenomena of Pellet-Cladding Interaction (PCI) and Pellet-Cladding Mechanical Interaction (PCMI) present significant challenges. These issues affect the design and safety of NPPs, mainly due to factors like fission gas release and fuel swelling. Accurately modeling these interactions is complex, as they are complexly linked with the fuel’s neutronic and thermal responses. Traditionally, the simulation of these interactions, essential for nuclear safety and risk assessment, has been complex and resource intensive. This study focuses on developing innovative surrogate models to enhance computational efficiency in nuclear fuel code analysis. A surrogate model based on a neural network approach is a data-driven computational model that approximates the behavior or output of complex, time-consuming, or resource-intensive simulations. This study specifically employs Artificial Neural Networks (ANNs) and statistical algorithms, aiming to reduce high computational cost of traditional approaches. A validated synthetic dataset, representing thermal analysis under steady-state conditions, is used to train the machine learning model. The dataset specifically focuses on Cladding Temperature. The study investigates eight different Test-Case. The surrogate model, trained on only 20% of the dataset, can predicts the entire time series of temperature using ARIMA, LSTM, and Prophet algorithms. The maximum error achieved by the surrogate model is 3.19°C compared to the validated temperature. The study demonstrates that surrogate models offer a time-efficient alternative for simulating complex physical phenomena, achieving a balance between accuracy and efficiency. This approach is particularly beneficial in scenarios where full-scale 2D and 3D simulations are excessively time-consuming, providing quicker results and significantly reducing computational resources.