Potential of MLP neural networks for gamma dosimetry using spectroscopy data

Fricke gel dosimeters are suitable choices for measuring gamma dose distribution using spectroscopy data. Repeating dose measurements for too many conditions under different circumstances is a costly time-consuming labor, that is not practically feasible. However, based on an amount of measurement data, an artificial neural network can predict the response of the dosimeters under a variety of circumstances. This eliminates the need to calibrate every batch. Potential of neural networks to predict the response of Fricke gel dosimeters was investigated using spectroscopy data in this work. A feed-forward multilayer neural network model was designed using MATLAB software. A 6x4x1 architecture was trained using the measured optical absorbance of the prepared dosimeter samples with different compositions irradiated by doses ranging from 10 to 60 Gy. Through the network, a prediction of optical absorbance was made, resulting in a mean squared error of 0.003 for the test data. The best validation mean squared errors is equal to 0.008 at the epoch 16. The output data has a strong correlation with the target values as indicated by the Pearson correlation coefficient, which is 0.991. Despite the limited dataset, it was demonstrated that the artificial neural network method performed well for determining the optical absorbance and, in turn, the gel dose response.