Quantitative Analysis Model for Cadmium Based on Energy-Dispersive X-ray Fluorescence and Convolutional Neural Networks

Energy-Dispersive X-ray Fluorescence (EDXRF) faces challenges in trace cadmium (Cd) quantification, including high interference and low sensitivity, which affect detection accuracy. Convolutional Neural Networks (CNNs), known for their strong feature extraction capabilities, offer a promising approach to improve EDXRF's performance in quantitative analysis. This study explores the effectiveness of a 1D Convolutional Neural Network (1DCNN) in predicting Cd concentrations in soybean X-ray fluorescence spectra, with a focus on how convolutional layer configuration and hyperparameters influence model convergence. We tested 16 different scenarios, varying the number of convolutional layers and filters per layer, and identified the optimal architecture as four convolutional layers with 16, 16, 32, and 64 filters. With a learning rate of 0.01 and 80 training epochs, the model achieved a 92% prediction accuracy on unseen soybean spectral data. These results demonstrate that 1DCNNs can effectively process EDXRF data, providing a novel method for Cd quantification.