Detection of cesium in salt-lake brine using laser-induced breakdown spectroscopy combined with convolutional neural network

Abstract

To meet the application needs for cesium (Cs) extraction from salt-lake brines, the present work explores a laser-induced breakdown spectroscopy (LIBS) method that facilitates sample analysis by breakdown near the liquid-air interface. This approach addresses the demand for in-situ analysis with a low detection limit and a wide detection range. Experimental studies were conducted using 14 samples with different concentrations (10-1000 ppm) prepared by adding various amounts of Cs into raw salt-lake brines. Utilizing a LIBS setup equipped with a high-speed camera, over 4200 sets of spectral data were obtained. The effects of focal offset on liquid disturbance and LIBS signal quality were studied in detail, and it was found that the optimization of the focal offset not only can suppresses the liquid disturbance, but also improves the signal quality, including signal-to-noise ratio, signal-to-background ratio. These findings are critical for the advancement of long-term, continuous, in-situ LIBS detection technology. To achieve precise Cs detection across a wide concentration range, two multivariate models were constructed based on convolutional neural network (CNN) with different input data (OD-CNN model with original data and AD-CNN model with augmented data). Both models were capable of Cs detection across a wide concentration range, and comparative studies demonstrated that the AD-CNN model outperforms the OD-CNN model. Specifically, the coefficient of determination value improved from 97.19% to 99.81% with the AD-CNN model, while the mean absolute error and root mean square error were reduced by 56.95% and 53.63%, respectively, compared to the OD-CNN model. These results highlight that the AD-CNN model provides a robust approach to mitigate the influence of matrix effects, making it suitable for in-situ LIBS monitoring during the process of Cs extraction from salt-lake brine.