Thumbnail Image
Publication

Accelerating Monte Carlo and deterministic neutron transport models using machine learning

Berry, Jessica J.
Citations
Altmetric:
Advisor
Osborne, Andrew
Editor
Date
Date Issued
2024
Date Submitted
Keywords
machine learning
 neutron transport models
 nuclear engineering
Collections
2024 - Mines Theses & Dissertations
Show all metadata
Files
Thumbnail Image
Berry_mines_0052E_12901.pdf
Adobe PDF, 6.23 MB
Thumbnail Image
Berry_mines_0052E_316/GZ_Permission.pdf
Adobe PDF, 188.7 KB
Thumbnail Image
Berry_mines_0052E_316/Romano_Permission.pdf
Adobe PDF, 209.67 KB
Thumbnail Image
Berry_mines_0052E_316/Permission_upsampling_LWR.pdf
Adobe PDF, 147.47 KB
Thumbnail Image
Berry_mines_0052E_316/Permission_first_paper.pdf
Adobe PDF, 276.12 KB
Research Projects
Organizational Units
Journal Issue
URI
https://hdl.handle.net/11124/179220
Embargo Expires
Abstract
Modeling and simulation are essential in the design, analysis and operation of nuclear reactors. These systems can be simulated using deterministic or Monte Carlo (MC) methods, with arbitrary fidelity given enough computational resources. Although maximizing accuracy and resolution is a priority in the modeling and simulation of reactor systems, achieving high fidelity is computationally intensive. Machine learning methods can dramatically improve the fidelity of nuclear reactor modeling and simulation with minimal additional computational cost. In this dissertation, the development and application of feed forward artificial neural network (ANN) and convolutional neural network (CNN) models for accelerating nuclear reactor simulation codes are presented. The ANN is used as a surrogate model for a deterministic multigroup neutron transport model. The ANN runs orders of magnitude faster than the multigroup model and can accurately pre-dict the latter’s sensitivity to key input parameters. The CNN models are trained to increase the spatial and energy resolution of quantities calculated by neutron MC simulations. Increasing the resolution of spatial and energy grids allows simulations of nuclear systems in more detail but this can have a sig-nificant impact on the computational resources required. The CNNs described in this dissertation can double the spatial and energy resolution of reactor MC flux tallies in each dimension with minimal additional computational cost. Importantly, the CNN models can predict high-resolution flux values with the same uncertainty as full MC simulations done at the equivalent resolution.
Associated Publications
Rights
Copyright of the original work is retained by the author.
Embedded videos