Application of Soft Computing Techniques in Modelling of Soaked and Unsoaked California Bearing Ratio

In this study, author attempted to establish a correlation between soil physical parameters and California Bearing Ratio of lateritic soils using advanced mathematical techniques such as the Support Vector Machine (SVM), Random Forest (RF), M5 tree, multiple linear regression, and Artificial Neural Network. A total of 480 soil samples were collected and separated into a data set using training and validation of the generated models based on the main soil parameters of Liquid Limit (LL), Plastic Limit (PL), Natural moisture content (NMC), Specific gravity (GS), Fines (F), Gravel, and Sand. The Principal Component Analysis (PCA) was used to minimize the dataset's huge dimension, and the approximate sum of the first four principal components (PC) captured 88 percent of the variability in the response variable with just 12% information loss. The RMSE values of 21.6, 21.23, 295.67, 7.03, 14.54 and 24.43,24.59,326.49,8.63,17.71 are from the MLR, ANN, MS Tree, RF, and SVM models for SCBR and USCBR values, respectively. For SCBR and USCBR, random forest (RF) yielded the lowest values of 7.03 and 8.63, respectively. Similarly, the R values range from 0.1 to 0.94 and 0.01 to 0.92, indicating that the anticipated and real SCBR and USCBR are related. The Random Forest Model for SCBR and USCBR was shown to be the best by the correlation coefficient values, while the MS tree model for SCBR and USCBR was determined to have the lowest coefficient of determination R2. As a result, it can be concluded that Random Forest provided the best Soaked and Unsoaked CBR model based on the dataset, while MS tree provided the poorest model. The model is a valuable tool for evaluating the subsurface indices of a civil engineering site at the preliminary planning stage before final structural design for the substructures, as the anticipated soil parameter values are within permitted accuracy.