Abstract:
California bearing ratio (CBR) test is one of the comprehensive tests used for the last few decades to design the pavement thickness of roadways, railways and airport runways. Laboratory-performed CBR test is considerably rigorous and time-taking. In a quest for an alternative solution, this study utilizes novel computational approaches, including the kernel ridges regression, K-nearest neighbor and Gaussian process regression (GPR), to predict the soaked CBR value of soils. A vast quantity of 1011 in situ soil samples were collected from an ongoing highway project work site. Two data divisional approaches, i.e., K-Fold and fuzzy c-means (FCM) clustering, were used to separate the dataset into training and testing subsets. Apart from the numerous statistical performance measurement indices, ranking and overfitting analysis were used to identify the best-fitted CBR prediction model. Additionally, the literature models were also tried to validate through present study datasets. From the results of Pearson’s correlation analysis, Sand, Fine Content, Plastic Limit, Plasticity Index, Maximum Dry Density and Optimum Moisture Content were found to be most influencing input parameters in developing the soaked CBR of fine-grained plastic soils. Experimental results also establish the proficiency of the GPR model developed through FCM and K-Fold data division approaches. The K-Fold data division approach was found to be helpful in removing the overfitting of the models. Furthermore, the predictive ability of any model is considerably influenced by the geological location of the soils/materials used for the model development.
