Development and Performance Evaluation of RAPID-AI for High-Fidelity Whole-Core Reactor Simulation: Analysis of Spatial Undersampling in Monte Carlo Eigenvalue Calculations

This paper evaluates the accuracy and computational performance of the RAPID (real-time analysis for particle transport and in-situ detection) code system and its newly developed physics-informed artificial intelligence (AI) extension, RAPID-AI, for whole-core reactor simulations. RAPID is based on the multistage response function transport (MRT) methodology, a hybrid deterministic Monte Carlo approach that combines the accuracy of high-fidelity transport methods with computational efficiency.The Watts Bar Unit 1 benchmark is used to assess the accuracy and performance of RAPID and RAPID-AI relative to the Serpent Monte Carlo code system. Comparisons were performed for the effective multiplication factor [k-effective (keff)] and the three-dimensional fission neutron distribution. The results demonstrate that RAPID can calculate the pinwise, axially resolved (144 axial intervals) fission neutron density distribution for the full reactor core in approximately 7 min using a single CPU core, while RAPID-AI performs the same calculation in approximately 13 s. In contrast, Serpent requires approximately 8 days on 40 CPU cores to produce an assemblywise, axially resolved fission neutron distribution.Both RAPID and RAPID-AI show excellent agreement with Serpent in the prediction of keff. Furthermore, the study demonstrates that obtaining an accurate pinwise, axially resolved solution using conventional Monte Carlo methods requires prohibitively large computational resources.To investigate the loss of spatial detail associated with assemblywise Monte Carlo tallies, a detailed single-assembly study was performed. The results reveal that Monte Carlo calculations may report very low relative statistical uncertainties while still exhibiting significant spatial undersampling within the model.These findings demonstrate that RAPID-AI provides an accurate and highly efficient framework for large-scale reactor simulations. Its ability to rapidly generate high-fidelity, physics-based solutions makes it a promising tool for real-time reactor analysis, digital reactor applications, and the generation of large training data sets for advanced AI/machine learning models in nuclear engineering.