Estimating grain size of microgranular material using laser-induced breakdown spectroscopy combined with machine learning algorithms

Recent work validated a new method for estimating grain size of microgranular materials in the range of tens-to-hundreds micrometers using laser-induced breakdown spectroscopy (LIBS). In that situation, univariate analysis was performed and a piecewise model has to be constructed for achieving the estimation of the grain size within such a wide range. This is due to the fact that a complex dependence of plasma formation environment (i.e., the status of luminous plasma and therefore LIBS signal to be measured) on grain size occurs in the size range studied there. In the present work, we tentatively construct a unified calibration model suitable for LIBS-based estimation of those grain sizes. Specifically, two unified multivariate calibration models are constructed based on back-propagation neural network (BPNN) algorithms using the feature selection strategies with and without considering physical prior knowledge, respectively. By detailed analysis of the performances of the two multivariate models, it was found that, a unified calibration model can be constructed successfully based on BPNN algorithms for estimating the grain size in the range of tens-to-hundreds micrometers. It was also found that this model constructed with a physics-guided feature selection strategy has better prediction performances. This study has practical significance in developing the technology for material analysis using LIBS, especially in the case that LIBS signal exhibits a complex dependence on the material parameter to be estimated.