Error Analysis of a Hybrid Control Drum Worth Model

This paper presents a perturbation-based model for control drum worth prediction which employs both physics-based and statistics-based components. The model can be expensive to create due to the requirement for full-core Monte Carlo eigenvalue calculations. Therefore, it is important to analyze how the errors in Monte Carlo calculated k_eff used for model training affect model performance. It was found that the error in predicted criticalities could average to 70 pcm in the most complex form of the model and 215 pcm in the simplest form of the model. Furthermore, it was found that the Monte Carlo uncertainty in quantities calculated with Serpent used to train the models had minimal impact on the error observed from the model. Lastly, one of the forms of the hybrid model could be trained in considerably less computational time if the Monte Carlo calculations were run to higher uncertainty in k_eff with a small penalty to model performance.