Fusing Raman Spectra with Fiber Metrics in Machine Learning Models That Predict the Physical-Mechanical Properties of...

Wood pulp manufacturers regularly monitor specific sets of physical and mechanical properties of production pulps to ensure that they meet paper-maker product-quality targets. Conventional methods based on time-consuming manual procedures serve as cornerstones of pulp quality control. However, low-throughput laboratory analysis presents substantial personnel and material resource demands. As a result, conventional methods cannot support real-time process management. Here, we describe an alternative approach that applies data fusion to combine two instrumental measures of pulp quality and provide instantaneous input for analysis by advanced machine learning models that predict 20 paper properties. This strategy uses Raman spectroscopy to capture molecular vibrational structure indicative of the chemical composition and morphological state of a pulp sample, together with pulp and fiber physical attributes such as bulk freeness and distributions of individual length, curl, and width measured online by a PulpEye integrated analysis system. The present work has collected these instrumental measures to build an X matrix of predictors for more than 500 production pulp samples. An industry-standard battery of conventional physical and mechanical quality control tests provides a complete set of response matrices, Y, that support the training of machine learning models that predict conventional measures of paper quality from instantaneous spectroscopic and physical attributes of pulp. These machine-learning algorithms emphasize features from both Raman spectra and PulpEye metrics. Among multivariate analysis methods, we find that eXtreme Gradient Boosting (XGBoost) algorithms consistently outperform conventional Partial Least-Squares Regression (PLSR). The best-trained models predict many target paper properties with errors comparable to those of wet-lab measurements, promising a method for real-time process control in the pulp and paper industry. Implementation will offer data-driven insights that reduce the reliance on manual testing and facilitate a deeper understanding of the factors that affect the properties of complex cellulosic materials.