International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering
Abstract
This study presents the ongoing development of a machine learning model as a diagnostic tool to identify particle cluster(s) formation during Monte Carlo criticality simulations. The particle distribution is essentially spatial-temporal features—evolves in space and simulation cycle—and can be directly extracted from Monte Carlo simulation for machine-learning model training without the use of intermediary/silhouette metrics—potentially losing information. The machine learning model uses a combination of convolutional neural network and long short-term memory (Many-to-One variant) to map the spatial-temporal features into higher time-independent features for the classification task. The model was trained with synthetic dataset generated by 3D Gauss distribution to mimic particle cluster(s) formation, and later evaluated with the test set generated from OpenMC simulations. The model achieves an accuracy of 99% for the synthetic test set but only 11% for the OpenMC dataset, revealing the limitations of current features extraction method. Furthermore, the direct use of particle distribution as spatial-temporal features is not viable for a large-scale machine learning model due to the prohibitively large dataset space requirement, so potential spatial-temporal features are discussed for further study.
Publisher
Canadian Nuclear Society, American Nuclear Society