Convolutional neural network use in fast neutron computed tomography for void fraction measurements

• A convolutional neural network was used to enhance fast neutron computed tomography image reconstructions. • Improvements in image reconstruction quality resulted in more accurate subchannel void fraction predictions. • Demonstrated noise and artifact reduction in synthetic CANDU nuclear fuel bundle phantoms. • Preprocessing methods such as an increase of field of view, reduced an overprediction bias. • Results support a method to obtain more accurate void fraction data for thermal–hydraulic safety analysis. Subchannel analysis codes are used for safety analysis and require suitable data to accurately model complex conditions that occur in fuel bundle geometries. To ensure accurate modelling of the complex two-phase flow phenomena in a reactor bundle, experimental data of the void distribution are required. Computed tomography provides a non-invasive method of measurement of the 2D or 3D distribution of void fraction within a bundle geometry. However, thick pressure vessels in full-scale conditions make it difficult to image bundle geometries and maintain contrast of the internal structures such as water/vapour when using photon-based sources. Fast neutron systems provide good penetration capability while maintaining sensitivity to the water-vapour contrast within the bundle but require long scan times (order of hours) because of the relatively poor detection efficiency and low source strength, which may preclude the application in full-scale thermal–hydraulic testing scenarios. Therefore, to effectively use Fast Neutron Computed Tomography (FNCT) in thermal-hydraulics safety experiments, the large scan times must be reduced, and the increased noise and decreased image fidelity that accompanies this reduction must be addressed. This challenge is addressed in this paper using convolutional neural networks (CNNs), a branch of machine learning that excels in image processing tasks. A Synthetic nuclear fuel bundle image reconstructions training dataset, including noise and blurring effects, was generated using a custom MATLAB and Python. The CNN model uses the dataset to create a mapping between these noisy reconstructed images and their respective ground truth image. In this work, the Residual U-Net model significantly improves image reconstructions, leading to more accurate measurements of subchannel void fraction compared to the initial noisy images. The model is able to predict the subchannel void fraction with a mean absolute percentage error (MAPE) of 3.2 % ± 2.4 % on a subchannel basis. Here, MAPE refers to the mean of the absolute differences between the predicted and true void fractions, expressed in percentage points. This means that a void fraction prediction of 50 % with a true value of 53 % results in a 3 % error, not a relative percentage of the void fraction itself. The void fractions were predicted within 8.43 % of the true values for 95 % of the subchannels, demonstrating that the majority of predictions are highly accurate.