Estimation of External Dose Volume Correction Factors Based on Neural Network

Junjun Chen,Jingyuan Qu,Junjun Gong
DOI: https://doi.org/10.1115/icone26-81616
2018-01-01
Abstract:Before the design of the nuclear facility, it is necessary to estimate the dose caused by radioactive material released into the environment during the course of the serious accidents. Semi-infinite hemisphere geometric model is established to estimate the personal external exposure dose outdoors in which the distribution of the nuclides is assumed to be uniform. The exposed staffs are in a limited cubic space when using this model to evaluate the controllability of the main control room. Thus, the volume correction factor is needed to correct the dose, whose traditional expression is f = 352/pow(V, 0.338). The formula cannot satisfy the requirement of higher accuracy due to the neglect of the influence of the shape of geometric model and γ-rays energy. Usually the actual control room is a cube and the γ-rays energies emitted from various nuclides are different. In order to calculate the accurate volume correction factor of main control room under different geometric conditions, a finite cubic geometric model is established in this paper. The length and width of the model are between 6m and 50 m, the height is between 4m and 6m, and γ-ray energy respectively are 0.05, 0.2, 0.733, 1.2 and 3 MeV, respectively. The effective volume values for different conditions are calculated by the Monte-Carlo program, and 318 groups of results are obtained. The calculated volume dose rate of 360m × 360m × 255m (assuming semi-infinite) cube at 733keV γ-rays energy is taken as a criterion, whose ratio of the other calculation results is the new volume correction factor value. By comparing two volume correction factors, the relative discrepancies are within 3 folds, proving that the calculation result is reasonable and feasible. The new volume correction factor varies with γ-rays energy and the shape of the geometric model. A neural network model corresponding to the volume correction factor is developed to apply to more cases. 80% of the results are randomly selected as the training set of neural network. The remaining 20% of the result as the test set of the cross-test is to predict the results of the trained neural network, whose relative errors are less than 5%. The neural network model can obtain the volume correction factor under different geometry and γ-ray energy conditions. Finally, a volume correction factor library is established, which can provide a powerful reference to obtain the volume correction factor of the limited space model such as the main control room.
What problem does this paper attempt to address?