Assessment and predicting the effect of accident tolerant fuel composition and geometry on neutronic and safety parameters in small modular reactors via artificial neural network and adaptive neuro-fuzzy inference system

This study investigates the application of Artificial Intelligence in nuclear reactors, focusing on the impact of Accident Tolerant Fuel (ATF) composition and geometry on Small Modular Reactors (SMRs) parameters. Leveraging Artificial Neural Networks (ANNs) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS), the research comprehensively examines the effects of cladding material (FeCrAl) modifications and burnable absorber concentration variations on key characteristics of the NuScale reactor. Neutronic calculations were meticulously conducted using MCNP6, a state-of-the-art Monte Carlo particle transport code, to assess reactivity, radial power peaking factor, feedback coefficients, and delayed neutron fraction. The results demonstrate that cladding thickness, chromium content, aluminum content, and gadolinia concentration significantly influence neutronic parameters. Furthermore, the study reveals intricate relationships between these parameters and reactor performance, providing valuable insights for reactor design and optimization. In addition to the aforementioned case studies and simulations, ANNs, and ANFIS were developed to predict key neutronic and safety parameters in the NuScale SMR loaded with ATF. The models, trained on extensive neutronic data, accurately predicted these parameters. The model's inputs included gadolinium concentration, cladding material weight percentage, and cladding thickness, while outputs encompassed excess reactivity, hot full power reactivity, effective delayed neutron fraction, radial power peaking factor, and fuel and coolant reactivity feedback coefficients. Both ANN and ANFIS models demonstrated exceptional accuracy and generalizability, offering a valuable tool for predicting the influence of ATF variations on reactor behavior. However, the ANN model consistently outperformed the ANFIS model, exhibiting lower prediction errors and demonstrating superior suitability for the intended application.