A machine learning approach for the prediction of cellulose nanofibril films’ mechanical properties from suspension morphological data
Published: 2025
Publication Name: Cellulose
Publication URL: https://link.springer.com/article/10.1007/s10570-025-06631-7
Abstract:
Cellulose nanofibrils (CNFs) are promising bio-based replacements for fossil-fuel based products with outstanding mechanical and barrier properties. The mechanical performance of CNF films is highly affected by CNF morphological features, such as length and width. To have fast and efficient production of CNFs and control over their final properties, it is necessary to explore the correlation between suspension morphological features and their final properties of the films. For this purpose, in this study, the effect of various CNF morphological features generated by a compact fiber analyzer on tensile strength, tensile modulus, and toughness of the produced films was explored by using machine learning (ML) models including linear ridge regression (LRR), random forest (RF), and gradient boosting machine (GBM) trained and validated on seven training samples. The number of features was reduced through an innovative feature engineering approach, and new features with high correlation with outputs were created. This reduction in the number of features reduced the complexity of the training process. By testing the trained ML models on two unseen data sets, it was concluded that LRR and RF models performed well in predicting tensile strength and toughness with a normalized root mean squared error (NRMSE) of 0.1–0.26. Finally, a sensitivity analysis was done using partial dependence plots (PDPs) to evaluate the impact of different features on the predictions made by the trained ML models. Sensitivity analysis revealed that the captured correlations between the input and output features were aligned with the existing literature, and it helped understand the underlying physics. The machine learning codes generated are shared publicly and can be freely used to predict the tensile properties of CNF films from the morphological properties of CNF suspensions.