The synergic approach between machine learning, chemometrics, and NIR hyperspectral imagery for a real-time, reliable, and accurate prediction of mass loss in cement samples.

Alternative and non-destructive analytical methods that predict analyte concentration accurately and immediately in a specific matrix are becoming vital in the analytical chemistry domain. Here, a new innovative and rapid method of predicting mass loss of cement samples based on a combination of Machine Learning (ML) and the emerging technique called Hyperspectral Imaging (HSI) is presented. The method has proved its reliability and accuracy by providing a predictive ML model, with satisfactory best validation scores recorded using partial least squared regression, with a reported ratio of performance to inter-quartile distance and root mean squared error of 12,89 and 0.337, respectively. Moreover, the possibility of optimizing and boosting the performance of the method by optimizing the predictive model performance has been suggested. Therefore, a features selection approach was conducted to disqualify non-relevant wavelengths and stress only relevant ones in order to make them the only contributors to a final optimized model. The best selected features subset was composed of 28 wavelengths out of 121, found by applying genetic algorithm combined to partial least squares regression as a feature selection method, on spectra preprocessed consecutively by the first-order savitzky-golay derivative calculated with 7-point quadratic SG filter, and multiplicative scatter correction method. The overall results show the possibility of combining HSI and ML for fast monitoring of water content in cement samples.