Validation of the neural network for 3D photon radiation field reconstruction under various source distributions

Introduction: This paper proposes a five-layer fully connected neural network for predicting radiation parameters in a radiation space based on detector readings.

Methods: The network is trained and tested using gamma flux values from individual detector positions as input, and is used to predict the gamma radiation field in 3D space under different source term distributions. The method is evaluated using the mean percentage change error (PCT) for the test set under different source term distributions.

Results: The results show that the neural network method can accurately predict radiation parameters with an average PCT error range of 0.53% to 3.11%, within the given measurement input error range of ± 10%. The method also demonstrates its ability to directly reconstruct the 3D radiation field with some simple source terms.

Discussion: The proposed method has practical value in real operations within radiation spaces, and can be used to improve the accuracy and efficiency of predicting radiation parameters. Further research could explore the use of more complex source term distributions and the integration of other types of sensors for improved accuracy.