A comparison between geomembrane-sand tests and machine learning predictions

Abstract

The interaction between soils and geosynthetics plays an important role in the applications of these materials for reinforcement in geotechnical engineering. The complexities of soil-geosynthetic interactions vary depending on the type and properties of both the geosynthetic and the soil. This paper introduces a machine learning approach, specifically a random forest algorithm, for predicting interface friction angles. The dataset comprises 495 interfaces involving geomembranes and sand, with 14 influencing parameters recorded for each interface, influencing the shear strength outcome. In the analysis, Pearson's correlation coefficient is employed to measure the linear interdependence between each pair of input-input and input-output variables. Following the linear regression analysis, an optimized random forest is utilized to project the interface friction angle. The random forest algorithm divides the selected data into training and testing sets, and only 3% of the training set and 6% of the testing set exceed ±5° from the actual records. The coefficient of determination (R2) indicates strong agreement between the predicted and laboratory study friction angles, with R2 = 0.93 for the training set and R2 = 0.92 for the testing set. Consequently, the random forest algorithm demonstrates effectiveness in predicting interface friction angles.