Machine learning-assisted spectrometric method for pulp extractive analysis based on model pulps

Abstract Extractives in wood, despite being vital in plant survival in adverse environmental conditions, can cause issues in pulping and papermaking, leading to significant financial losses within the forest industry. Conventional analysis is time consuming due to the challenging sample preparation, data interpretation and complex nature of extractives. There is a clear need for the development of a quick, non-destructive method for on-line pulp extractives monitoring. In this work, a machine learning-supported procedure for classification and prediction of extractives based on near infrared (NIR) and Raman spectroscopies was proposed. To avoid the influence of many variables, the method was developed and validated using a model compound approach, where cellulose was spiked with model extractives compounds. The accuracy of sample classification depending on the extractive added was 92.4 % based on NIR. The accuracy of classification of six samples containing different concentrations of the model compound using NIR data was 89.5 %. Partial least squares calibration model applied to pretreated NIR spectra yielded R 2 and root standard error of 0.78 and 0.35, respectively This means that the method could be used for non-selective quick estimation of extractive content with potential application to forest industry in process and fiber quality control in pulp and paper.