Determination of low Z elements concentrations in geological samples by energy dispersive X-ray fluorescence with a back propagation neural network

Due to complex scattering from the sample dark matrix, absorption in the detector window and the competing Auger effect with higher cross-section for low Z elements (Z < 14), direct quantification of low Z elements by measuring characteristic X-ray fluorescence intensities during energy dispersive X-ray fluorescence (EDXRF) analysis of geological samples is challenging. This paper reports the chemometric quantitative analysis model of low Z elements (O, Na, Mg, and Al) in geological samples obtained by training the backpropagation neural network (BPNN) using the Compton scatter data combined with the concentrations of measurable elements. The training data is derived from the measured spectra of soil and rock standard samples and their synthetic samples configured with compounds. The standardized Compton scatter data and correlated element concentrations were used as the input data of the BPNN model. The prediction results of the BPNN model show that the coefficient of determination ( R 2 ) values between true and predicted concentrations for O, Na, Mg, and Al are both >0.95. It implied that the modeling approaches significantly overcome matrix effects between the concentrations of low Z elements and Compton scatter peaks. So, the method has the potential for being widely used in the analysis of samples rich in low Z elements. • A BPNN quantitative analysis model of low Z elements in geological samples was developed. • The Compton scatter data ware used for model training. • The model prediction performance is improved by training on correlated elements concentration. • The application of K-fold cross validation improves model accuracy.