Leveraging machine learning capabilities for the characterization of irradiated uranium : a case study of prediction methods for nuclear safeguards and nuclear forensics

The characterization of irradiated actinide materials is a complex multi-variate problem that is relevant in both nuclear forensics and nuclear safeguards. Compounding factors such as irradiation history, initial actinide composition, enrichment, decay time, neutron flux, and irradiation energy spectrum, many of which may be either poorly constrained or entirely unknown, make source attribution challenging. This work demonstrates the capabilities of machine learning to advance the state-of-the-art measurements that are essential to nuclear safeguards and forensics. Predictive models were constructed with regularized linear methods, decision trees, and ensembles of decision trees for simultaneous irradiation time (t[subscript irr]) and decay time (t[subscript dec]) independent determination of uranium enrichment via gamma emissions with synthetically generated data. The t[subscript dec] independent predictive enrichment monitor was then validated with real-world experimentally measured data while holding t[subscript irr] fixed. All models were trained and evaluated on datasets consisting of exactly seven discrete enrichment values (0.02%, 0.71%, 3%, 20%, 50%, 63%, and 97% ₂₃₅U). Models were trained to predict the enrichment of randomly selected gamma-ray emission profiles with varying t[subscript dec] values but with t[subscript irr] fixed at 1 hour. Additional t[subscript irr] values were introduced into the dataset and new models were then trained. Mismatches between the training and testing dataset t[subscript irr] values were introduced to characterize the generalization performance of models to data which was unrepresented during training. It is shown that for the ranges of t[subscript irr] considered, there is no statistically significant degradation in model performance, indicating robust generalization performance. The adaptively boosted decision tree ensemble constructed on these simulated data with fixed t[subscript irr] was able to perform predictions on uranium enrichment with a mean absolute error of 1.7% of the ₂₃₅U enrichment values for fixed t[subscript irr]. Experimental measurements were performed to validate the fixed t[subscript irr] modeling capability. An adaptively boosted decision tree ensemble constructed on the experimental data achieved a mean absolute error of only 0.3% of the ₂₃₅U enrichment values, outperforming the models trained on synthetic data. Next, due to performance limitations in computing power principal component analysis (PCA) was used for dataset feature truncation. It is shown that PCA results in a substantial reduction in computation time. Furthermore, PCA also improves the performance of models by removing unimportant variations in the data. A decision tree evaluated on PCA truncated data achieved a mean absolute error of only 0.05% of the ₂₃₅U enrichment values. The prediction capabilities provided by these models can be naturally extended to application-focused measurements in the fields of nuclear safeguards, nuclear forensics, and nuclear nonproliferation. The scenario presented here represents the first step towards building these application-focused systems