Development of an artificial neural network model for generating macroscopic cross-sections for RAST-AI

• Published in: Annals of nuclear energy 2023-06, Vol.186, p.109777, Article 109777
• Main Authors: Dzianisau, Siarhei, Saeju, Korawit, Lee, Hyun Chul, Lee, Deokjung
• Format: Article
• Language: English
• Subjects: Artificial neural network; Cross-section; Nuclear engineering; Pin power
• ISSN: 0306-4549
• DOI: 10.1016/j.anucene.2023.109777

•An artificial neural network model capable of generating macroscopic cross-sections and pin powers was developed.•An optimization method for fine-tuning variations of input parameters in training dataset was introduced.•The reduction in time required to generate the optimized training dataset exceeded 3 times.•The neural network model performance was greatly improved after using the optimized training dataset of a smaller size.•A new training, testing, validation dataset preparation philosophy was introduced and compared to conventional approach. Homogenized macroscopic cross-sections (XS) are necessary for running core-wise nodal diffusion calculations. XS sets are usually generated using time-consuming lattice physics codes. In this study, a pre-trained artificial neural network was developed and used for XS generation. The model was trained to produce macroscopic XS, pin powers, and assembly discontinuity factors for 16 × 16 and 17 × 17 fuel assembly types with independent variable enrichments of each fuel pin loaded with fresh UO2 fuel without burnable poisons. The training dataset optimization method was described and used for defining the required number of variations in input parameters, such as pin arrangements and thermal hydraulics parameters. The optimized dataset's generation took only 248 core-hours, which is below 3 days on a modern 4-core CPU. For the worst-case out-of-range testing data, the maximum observed difference with the reference was found below 3% for pin powers, and below 4.5% for XS values.