Deep learning-based spectrum-dose prediction for a plastic scintillation detector

Abstract

Dose-rate monitoring in workspaces plays an important role in protecting workers from radiation. Operational quantities are utilized to assess occupational exposures of workers. Typical detector materials for measuring these quantities include ion chambers or high-density materials. However, plastic scintillators are rarely used because of the absence of photo-peaks. We developed a deep learning-based method that can predict the ambient dose equivalent (H*(10))-a representative operational quantity-from measured spectra of plastic scintillation detectors to overcome their drawbacks in terms of spectroscopic dosimetry applications. To train the deep learning model, numerous gamma spectra with arbitrary energies of gamma rays and their H*(10) were calculated by Monte Carlo simulations and used as the dataset. Several neural network models were implemented by the Bayesian-based hyperparameter optimizations, and an ensemble model was used as the final model to enhance accuracy and generalization ability. The performance of the ensemble model was verified using simulated and measured spectra for representative radioisotopes. Furthermore, we confirmed that dose-rate prediction errors of the model were within acceptable uncertainty ranges suggested by the IAEA safety guide and that energy responses of the model satisfied IEC requirements.