Dynamic dose equivalent rate estimation in dismantling: Physic-informed surrogate modeling

Abstract

The estimation of dose equivalent rate plays a key role in the radiation protection strategy for decontamination and dismantling. In particular, real-time map estimation of dose equivalent rate provides a user interface for planning interventions on basic nuclear installations undergoing dismantling. Conventional approaches for this estimation rely either on Monte-Carlo simulation whose computational time is prohibitive for real-time applications or on deterministic approaches whose approximations deteriorate the precision of the estimation in complex configurations. This work focuses on the construction of surrogate models, designed to mitigate these limitations and to estimate in real-time the dose equivalent rate given a source position in a specific installation. These models are tuned using data from Monte Carlo simulations, and take advantage of additional information, called ”additional descriptors”. These descriptors embed the knowledge on the physical behavior of the particles in a specific simulated installation. Three surrogate models, the K nearest neighbors, XGBoost, and the Gaussian Process regression are compared, with and without the additional descriptors. They are evaluated on three configurations encountered in radiation protection. The results show that the physical information allows the surrogate models to adapt to new source positions in the geometry, and limits the size of the database needed to train the models.