ARABIAN JOURNAL OF GEOSCIENCES, cilt.15, sa.7, 2022 (SCI-Expanded)
Shear strength parameters are essential factors in many geotechnical problems. These parameters can be obtained by laboratory tests such as the direct shear test or triaxial test and can be interpreted from field tests such as the standard penetration test and the cone penetration test. All these approaches can be considered difficult and time-consuming, especially when dealing with soils containing large-sized particles. This study investigates the application of machine learning techniques to predict the peak friction angle of granular soils based on their index properties. For this purpose, a series of medium-scale direct shear tests were conducted for different soil specimens to construct the dataset needed to explore these techniques. Index properties such as soil classification, effective particle sizes, gradation coefficients, and relative density were used as inputs for the models. The techniques were employed are linear regression, M5 tree, random forest, K-nearest neighbors, locally weighted linear regression (LWL), and artificial neural networks (ANN). It is found that the random forest, the ANN, and the LWL models could reliably predict the peak friction angle of granular soils while the other models have achieved poorer results. Also, it is put forward that picking the proper technique and fine-tuning the model hyperparameters are essential aspects to consider when applying these techniques.