On the feasibility of machine-learning model for neutronic analyses of pressurized water reactor fuel assemblies

Neutronic analysis of nuclear fuel assemblies is crucial for optimizing reactor core design. The conventional techniques involve computationally expensive simulations. This work explores the feasibility of developing efficient machine-learning models for neutronic analysis. This was done by leveraging artificial neural networks (ANNs) for evaluating two important neutronic parameters: neutron multiplication factor (k ∞ ) and void coefficient (α v ), of 17 × 17 pressurized water reactor (PWR) fuel assemblies. OpenMC neutronics code was employed to generate a training set of 6000 data points for the machine learning models with six input parameters: U-235 enrichment, densities of fuel and moderator, concentration of burnable poisons in fuel and moderator, and fuel pitch. Results reveal that the four-layer ANN model with sigmoid activation functions in the hidden layers exhibited excellent performance. The ANN predicts k ∞ with an average predictive error below 106 pcm, well within the acceptable limit of 500 pcm. Among the optimization algorithms employed, the simplex search method outperformed genetic algorithms, simulated annealing and particle swarm optimization. Utilizing the developed ANN model in simplex search, the correlation was derived for fuel pitch separating the under- and over-moderated regions under cold zero power (CZP) conditions. It was observed that the fuel pitch must be kept ≤ 1.434 cm to maintain the reactor’s under-moderated status throughout its operational lifetime. Overall, the accuracy of ANN model was found to be satisfactory, opening doors for further improvements.