Using machine learning to predict paperboard properties - a case study

This study aims to investigate the applicability and accuracy of different machine learning (ML) methods for predicting paperboard properties based on raw material and process data. The goal is not to find an ML method that can predict all properties simultaneously, but rather the most suitable method for a single property. The examined properties are bending stiffness and the curl. Furthermore, the focus is not on the most accurate prediction model, but on the best model to interpret. The results show that ML methods are applicable for predicting paperboard properties based on raw material and process data, e. g., bending stiffness and curl in machine direction (MD) and cross-machine direction (CD). The feature selection is crucial for the quality of the prediction. Therefore, domain knowledge is needed to select the right features. This selection improves the model and ensures its reliability, comprehensibility, and interpretability. The self-implemented recursive feature addition (RFA) takes the features specified by a domain expert for the feature selection into consideration. It requires less computation time in comparison to the common recursive feature elimination (RFE). The extremely randomized tree (ET) was chosen due to its high interpretability and good results. It achieves a mean absolute error (MAE) for the bending stiffness MD of 0.913 mN m; an MAE for curl MD of 0.058 m(-1) (in a value range from 7 mN m to 100 mN m and -1 m(-1) to 1 m(-1), respectively); 0.474 mN m and 0.163 m(-1) (in a value range from 5 mN m to 45 mN m and -1 m(-1) to 2 m(-1), respectively) for the respective CD properties.