Interpreting the relationship between properties of wood and pulping & paper via machine learning algorithms combined with SHAP analysis

Abstract The pulping ability and quality of paper high relay on the wood properties. However, the relationship between them are profound. Based on the extracting digital information from the anatomical, chemical, and physical properties of poplar wood, predictive models were developed for paper properties (tensile index, burst index and tear index) and pulping properties (Kappa number and pulp yield) using six algorithms, namely PLSR, ENR, RF, XGBoost, LightGBM, and CatBoost. The prediction results revealed that among the six algorithms, PLSR, ENR, and RF exhibited results of most prediction greater than 0.79. Notably, XGBoost, LightGBM, and CatBoost algorithms demonstrated superior predictive performance, with results greater than 0.9, except for the tear index. Furthermore, SHAP analysis suggested that the cellulose content is the primary factors to modulate pulping ability and the morphological features of cell wall shows apparent effects on mechanical properties of paper. It hopes the result will benefit to provide information to evaluate the value of poplar wood from different resources and then deliver instructions to genetic breeding program and forest management of poplar plantation.