Application of machine learning methods for spent fuel characterization based on gamma spectrometry measurements

Machine learning models have been developed to predict properties of spent nuclear fuel assemblies from Paks NPP, such as burnup, cooling time, initial enrichment, and Pu-239 content. Measured gamma spectra are processed, and activity ratios of fission products are calculated to serve as input features for Support Vector Regression, Random Forest, and Multi-Layer Perceptron models. Data uncertainty is considered during inference to produce prediction intervals, and input features are ranked using the Gini importance of Random Forest models. A deep learning approach using Convolutional Neural Networks has also been developed to predict the spent fuel parameters from the measured spectra directly. The new models can predict spent fuel parameters with great precision, outperforming earlier approaches that rely on nonlinear regression using a single feature to predict burnup and cooling time and can estimate initial enrichment and Pu-239 content.