Prediction of California Bearing Ratio Using Soil Index Properties by Regression and Machine-Learning Techniques

This study proposes regression and machine-learning techniques to develop a validated model that predicts the California Bearing Ratio (CBR) values for subgrade soil using soil index properties. Around 60 specimens were prepared experimentally by adding different sand percentages to the natural soil to provide a wide range of soil properties. In addition, soil test reports from the local transportation authority were also used in the study. A total of 110 soil samples were included to generalize the predicted model. This study included three machine-learning (ML) techniques: artificial neural networks (ANN), M5P Model tree, and the lazy algorithm K-nearest neighbor. In addition, two conventional modeling techniques were used: multiple linear regression (MLR) and nonlinear regression (NLR). In the developed model, the laboratory-determined CBR represents the response variables, whereas the compaction characteristics (optimum moisture content (OMC) and maximum dry density (MDD)), Atterberg limits (liquid limit (LL), plastic limit (PL), and plasticity index (PI)), density, gradation parameter (percent of materials retained on sieve #200 ( R 200 ), and percent of materials retained on sieve #10 ( R 10 )) were used as predictors. Results revealed that the best model to predict the CBR for soil using material properties is the ANN model with R 2 of 90.46 and RMSE of 7.89, followed by KNN, MLR, M5P, and nonlinear regression in descending order.