Machine Learning Methods for the Prediction of Abraded Butadiene-Styrene Rubber Surface Roughness

Abstract

The aim of this investigation was to evaluate the ability to predict the roughened surface area of soft polymeric materials using the machine learning method. Random forest and extremely random forest machine learning methods have been applied to predict roughened monolithic butadiene styrene rubber surface area based on the grain diameter of the abrasive paper. It was found that there is a very strong negative correlation between surface roughness and abrasive grain diameter (the Pearson correlation coefficient is equal to - 0.83 and the Spearman correlation coefficient is equal to - 0.91). The statistics of Shapiro-Wilk criteria confirmed that independently of the diameter of abrasive grain, the lengths of rubber surface profiles are normally distributed. It was determined that random forest and extremely random forest algorithms are suitable tools for predicting rubber surface area in dependence on abrasive grain diameter. The forest structure in which results obtained best coincidence with the observed (R2 = 0.95): for the random forest model the number of trees is equal to 20, the leaf size is equal to 4, at the subset of the feature (abrasive paper grain diameter) equal to 1. In the case of an extremely random forest, the number of trees is equal to 350, the size of the subset of the features is equal to 1, and the leaf size is 2. The experimental and generated results agree well. The relative approximation error does not exceed 0.079 %.

Keywords

• surface roughness
• prediction
• machine learning
• random forest
• extremely random forest