Machine learning for the prediction and optimization of production of cellulose nanocrystals by sulfuric acid hydrolysis

Developing an optimal method for cellulose nanocrystals (CNCs) production is promising and sustainable. Hence, machine learning (ML) algorithms were applied to aid the CNC preparation and optimization with the consideration of related factors in hydrolysis process. The dataset collected from published literatures were used to train the ML models for prediction and optimization of CNC production by sulfuric acid hydrolysis of different cellulose sources. The gradient boosting decision tree algorithm was the best one for the yield prediction (R2=0.86, RMSE=9.15), and for the crystallinity prediction (R2=0.87, RSME=2.56). The acid concentration and cellulose source were identified as the most important features of yield and crystallinity prediction, respectively. Shapley additive explanation is used to visually interpret the ML model and the interaction effect of input features on yield. Then, the ML models were optimized and evaluated by experimental validation. The predicted CNC yield is 61% at an acid concentration of 60%, ratio of acid/cellulose of 8, temperature of 54 °C, and time of 50 min. The optimized result was experimentally validated and the CNC yield of 58.3% with errors of less than 4.6%. This study provides new perspectives and opportunities to understand and improve the preparation of CNCs.