High-fidelity reactor modeling can generate significant volumes of data informed by low-uncertainty measurements that can be used to construct inverse machine learning models. Simulated environmental samples can be used with advanced data analytical methods to improve techniques for estimating reactor operating characteristics such as burnup. Models that incorporate additional sample properties like cooling time are expected to be more accurate than ones that train only on nuclide composition. Several approaches for estimating multiple sample properties from nuclide composition measurements have been developed and tested. These approaches fall into two broad categories: methods that aim to determine core-average burnup and cooling time simultaneously (either implicitly or explicitly) and methods that conduct serial estimation analyses. Numerous variations in analysis design, model specification, and training data construction were tested. For relatively equivalent volumes of training data, a model that directly estimates burnup independent of differences in cooling time (i.e., implicit cooling time estimation) was more accurate than an ensemble of burnup models at estimating burnup. Surprisingly, even though using cooling time as a feature for burnup estimation improves the performance of such an ensemble, the single model was still most effective because an ensemble is dependent on the precision of the cooling time estimator.
Year
2024
Abstract