The United States and Europe aim to convert research reactors from running on high-enriched uranium (HEU) to using low-enriched uranium (LEU) fuel, because maintaining a handling permit for HEU is costly and difficult. Despite that most fuel technologies require HEU, there is an alternative LEU fuel technology that can be used to run research reactors but has a very low tolerance for inhomogeneities in the uranium enrichment. Before this fuel can be produced cheaply, a technique for monitoring the uranium enrichment must be developed that is accurate, near-real-time and nondestructive.
Gamma spectrometry is a promising approach but does not reflect the uranium enrichment directly, relying instead on indirect methods which are not accurate enough to monitor the LEU fuel production. Multi-group analysis for uranium (MGAU) is one of these methods and can calculate the uranium enrichment from arbitrary sample geometries without the need to calibrate the measuring system’s geometry.
We utilize the MGAU principal and propose an artificial neural network (ANN) with a gamma spectrum as input and the corresponding uranium enrichment as output. A simulation of the physical processes in a uranium sample produced the training and test data.
I used different training sets and hyperparameter tuning with different loss functions to find an ANN model with an accuracy of 0.39 wt%. The reached accuracy is better than the accuracies Gunnink et al. reported originally (1-2 wt%). It is comparable with values for optimized versions of MGAU (0.2-0.8 wt%) but yet not accurate enough for the fuel production process (0.2 wt%).

Keywords:
Neural Networks, Uranium Enrichment