Utilizing BP neural networks to accurately reconstruct the tritium depth profile in materials for BIXS

Abstract: β-ray-induced X-ray spectroscopy (BIXS) is a promising method for tritium detection in solid materials because of its unique advantages, such as large detection depth, nondestructive testing capabilities, and low requirements for sample preparation. However, high-accuracy reconstruction of the tritium depth profile remains a significant challenge for this technique. In this study, a novel reconstruction method based on a backpropagation (BP) neural network algorithm that demonstrates high accuracy, broad applicability, and robust noise resistance is proposed. The average reconstruction error calculated using the BP network (8.0%) was much lower than that obtained using traditional numerical methods (26.5%). In addition, the BP method can accurately reconstruct BIX spectra of samples with an unknown range of tritium and exhibits wide applicability to spectra with various tritium distributions. Furthermore, the BP network demonstrates superior accuracy and stability compared to numerical methods when reconstructing the spectra, with a relative uncertainty ranging from 0 to 10%. This study highlights the advantages of BP networks in accurately reconstructing the tritium depth profile from BIXS and promotes their further application in tritium detection.