Prediction of cellulose micro/nanofiber aspect ratio and yield of nanofibrillation using machine learning techniques

Abstract

Predictive monitoring of two key properties of nanocellulose, aspect ratio and yield of nanofibrillation, would help manufacturers control and optimize production processes, given the uncertainty that still surrounds their influential factors. For that, 20 different types of cellulosic and lignocellulosic micro/nanofibers produced from spruce and pine softwoods, and by different pre-treatment and fibrillation techniques, were used as training and testing datasets aiming at the development and evaluation of three machine learning models. The models used were Random Forests (RF), Linear Regression (LR) and Artificial Neural Networks (ANN), broadening the scope of our previous work (Santos et al. in Cellulose 29:5609–5622, 2022. https://doi.org/10.1007/s10570-022-04631-5). Performance of these models were evaluated by comparing statistical parameters such as Mean Absolute Percentage Error (MAPE) and R². For the aspect ratio and the yield of nanofibrillation, inputs were chosen among these easily controlled or measured variables: Total lignin (wt%), Cellulose (wt%), Hemicellulose (wt%), Extractives (wt%), HPH Energy Consumption (kWh/kg), Cationic Demand (µeq/g), Transmittance at 600 nm and Consistency index (Ostwald-De Waele’s k). In both cases, the ANN models trained here provided satisfactory estimates of aspect ratio (MAPE = 4.54% and R2 = 0.96) and the yield of nanofibrillation (MAPE = 6.74% and R2 = 0.98), being able to capture the effect of the applied energy along the fibrillation process. RF and LR models resulted in correlation coefficients of 0.93 and 0.95, respectively, for aspect ratio, while for yield of nanofibrillation the correlation coefficients were 0.87 and 0.92.