| description abstract | Subgrade strength is characterized by many indices such as the California bearing ratio, resilient modulus, and dynamic modulus. The California bearing ratio (CBR) of the subgrade soil is an important input parameter in the design of flexible pavements and the CBR of soil is largely influenced by its basic and mechanical properties such as gradation, maximum dry density (MDD), optimum moisture content (OMC), liquid limit (LL), plastic limit (PL), and plasticity index (PI). In many cases, the existing statistical relations between CBR and basic soil properties do not show good accuracy as they are developed based on limited data points of specific soil types. It is imperative to use more sophisticated and advanced algorithms to predict the CBR of the soil. In recent years, artificial intelligence (AI)/machine learning techniques have gained popularity in geotechnical engineering in solving problems when deterministic solutions are unavailable or are very expensive in terms of computational cost. The present study uses a large amount of input data consisting of 679 data points from published literature and attempts to predict the CBR of soil using a K nearest neighbor (KNN) regressor, random forest regressor (RFR), support vector regressor (SVR), multilayer perceptron (MLP), and multilinear regression (MLR). The RFR is found to result in the maximum accuracy of predicted CBR followed by KNN, MLP, SVR, and MLR methods. Additionally, sensitivity analysis is performed and the results show that MDD followed by the percentage of gravels are the most influencing parameters affecting CBR compared to the percentage of fines and Atterberg limits (LL and PL) of the soil. Furthermore, the conventional curve-fitting approach is used to find the best-fit curve for the variation of CBR with MDD and the percentage of gravel (the two most sensitive parameters). The coefficient of determination of CBR based on curve fitting and the RFR model trained on the stated two features are compared. The RFR model is found to give more accurate predictions than the conventional regression analysis examined in this study. | |
