Artificial Neural Networks (ANNs) and Partial Least Squares (PLS) Regression in the Quantitative Analysis of Respirable Crystalline Silica by Fourier-Transform Infrared Spectroscopy (FTIR)

Abstract

Abstract Respirable crystalline silica (RCS) overexposure can lead to the development of silicosis which is a chronic, irreversible, potentially fatal respiratory disease. The most significant prerequisite for any silica exposure control plan is an accurate occupational exposure assessment. The results of crystalline silica analysis are often affected by other mineral interferences and are influenced by an analyst’s knowledge of mineralogy to accurately interpret infrared spectra and correct matrix interferences. Partial least squares (PLS) and artificial neural networks (ANNs) are two multivariate calibration methods to overcome the problem of spectral interferences without the need for an analyst intervention. The performance of these two methods in quantitative analysis of quartz in the presence of mineral interferences was evaluated and compared in this study. Fifty mixtures with different crystalline silica content ratios were prepared by mixing quartz with four common mineral interferences including kaolinite, albite, muscovite, and amorphous silica. Fourier-transform infrared spectra of the mixtures were split into training and test datasets. The optimal architecture of the ANN model was achieved using a two-level full factorial design experiment and data were modeled using ANN and PLS regression analysis. Root mean squared error of prediction values of 1.69 and 6.12 µg quartz for ANN and PLS models, respectively, revealed the fact that the both models performed very well in quantitative analysis of quartz in the presence of mineral interferences, with a better relative performance of the ANN model which can be related to the inherent nonlinear predictive ability of ANNs. Given the excellent predictive ability of the ANN model which can deal with a completely overlapped peak without any need of user’s intervention, it is recommended that the ANN model be optimized in future studies and utilized for reliable and rapid on-field assessment of RCS exposure.