LAUSR

Peak shear strength of soil-Geocomposite Drain Layer (GDL) interfaces is an important parameter in the designing and operating related engineering structures. In this paper, a database compiled from 316 large direct shear tests on soil-GDL interfaces has been established. Based on this database, five different machine learning models: Back Propagation Artificial Neural Network (BPANN) and Support Vector Machine (SVM), with hyperparameters optimised by Particle Swarm Optimisation Algorithm (PSO) and Genetic Algorithm (GA), respectively, and Extreme Learning Machine (ELM) optimised by Exhaustive Method, were adopt to assess the peak shear strength of soil-GDL interfaces. Then, a comprehensive investigation and comparison of the predictive performance for the models was conducted. Also, based on the selected optimal machine learning model, sensitivity analysis was conducted, and an empirical equation developed based on it. The research indicated that GA and PSO could significantly increase forecasting precision in a small number of iterations. The BPANN model optimised by PSO has the highest forecasting precision based on the statistics criteria: Root-Mean-Square Error, Correlation Coefficient, Coefficient of Determination, Wilmot’s Index of Agreement, and Mean Absolute Percentage Error. The normal stress has the biggest impact on the peak shear strength, followed by drainage core type, moisture saturation of the soil layer, shearing surface, soil type, consolidation condition, geotextile specification, soil density and drainage core thickness, and the ranking is affected partly by the data distribution of input parameters in the database based on mechanism analysis. An empirical equation developed from the optimal model was proposed to estimate the peak shear strength, which provides convenience for geotechnical engineering personnel with limited knowledge of machine learning technique.