Modeling of the mass attenuation coefficients of X ray beams using deep neural networks (DNN) and NIST database

Abstract

Attenuation coefficients are essential physical parameters for many applications, such as the calculation of photon penetration and energy deposition to evaluate biological shielding. Estimating these parameters is complex, making it necessary to apply more sophisticated methodologies. The objective of the present study was to propose a model for estimating the attenuation coefficients using artificial neural networks. The NIST database was used to estimate the attenuation coefficients in terms of energy and atomic number from a regression problem using two approaches: the proposition of an automated model using the framework Talos and a manual model using Keras. The characteristics of the best model proposed in Talos were applied in manual training via Keras with cross-validation to evaluate the learning curves. The following were also assessed: the comparison of the curves of the attenuation coefficients predicted by the model compared with the reference data and the general comparison of the vectors X and y of the two models discussed. The Talos framework reference model obtained the following values ​​of Loss and MSE error metric: 0.13 and 0.037, respectively. The best model of the manual approach received the following results: 0.19 and 0.08 for the loss function and MSE error metric, respectively. The absolute percentage error (MAE) of the difference in the results between the two models was: 0.065 and 0.044 for the Loss and MSE metrics. Despite applying two distinct propositions, both models had the same difficulties in predicting discontinuities in the physical behavior associated with the attenuation coefficients.

Keywords

• mass attenuation coefficient
• Neural Network
• deep learning