Detection of Localized Perturbations in Water-cooled Small Modular Reactors Using Machine Learning

Licentiate thesis, 2025

This thesis explores the application of machine learning methods for localizing
perturbations in water-cooled Small Modular Reactors (SMRs) using simulated
neutron noise data. The study addresses the challenge of identifying
perturbations in SMRs, where the dominance of the point-kinetic component —
containing no spatial information about the source — makes localization more
difficult. To investigate this, neutron noise behavior in large and small cores is
first compared, in which the characteristic contributions of thermal and fast
neutron groups in both cores are examined. Thermal noise is found to deviate
more from point-kinetic behavior, offering richer localization information.
A machine learning model for absorber of variable strength localization in a 2D
SMR geometry is trained on simulated relative thermal neutron noise data using
samples containing one or two sources. The model demonstrates a high
generalization capability by successfully localizing perturbations in samples with
up to ten noise sources. The study extends to localize perturbations using data
from sparse in-core detector configurations, where the model maintains
reasonable performance.
In the 3D scenario, the full neutron noise distribution of the Westinghouse
AP300 SMR model is used to train a model on samples with up to ten sources.
Different input representations of the complex-valued noise are considered. A
two-channel representation based on the real and imaginary parts, respectively,
of the neutron noise achieves improved metric scores in scenarios with more than
ten sources, demonstrating efficient information preservation with low input
dimensionality.
The results confirm the ability of machine learning models to localize
perturbation sources in complex 3D water-cooled SMR environments. This work
provides a foundation for data-driven reactor diagnostics and serves as a first
step towards the formulation of a complete framework to localize and classify
perturbations in SMRs using neutron noise.