Machine Learning Applications in Predicting Aluminum Material Thickness from Transmission X-ray Spectra: A Comparative Study of Feedforward Neural Network (FNN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) Models

Abstract

This study presents a machine-learning approach for predicting the thickness of aluminum plates using multi-energy X-ray transmission spectra. Three neural network architectures are investigated, including a Feedforward Neural Network (FNN), a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN). The experiment setup employs a ²⁴¹Am radioactive source with a total activity of approximately 1776 MBq to excite a composite target containing Zr, Sb, and Ba, thereby generating six characteristic X-ray lines at 15.78 keV (Zr–Kα), 17.67 keV (Zr–Kβ), 26.36 keV (Sb–Kα), 29.73 keV (Sb–Kβ), 32.2 keV (Ba–Kα), and 36.38 keV (Ba–Kβ). The collimated radiation beams are transmitted through aluminum samples of varying thicknesses, and the resulting spectra are recorded using a Si(Li) semiconductor detector. From the mesured spectral, the attenuation behaviour of X-ray intensity as a function of material thickness is extracted and used to train and evaluate the machine-learning models. The hyperparameters are subsequently optimized to obtain the most effective model configuration. After optimization, transmission spectra corresponding to aluminum samples of various thicknesses are input into the models to further validate their prediction capability. The results show that all three architectures - FNN, CNN, and RNN - achieve high predictive accuracy, with relative errors below 5% compared with the reference values. Such results confirm the effectiveness of integrating machine-learning techniques with multi-energy X-ray transmission for the quantitative determination of aluminum thickness and demonstrate the potential of this approach for practical non-destructive testing applications.