Prediction of California Bearing Ratio of Fine Grain Soil from Index Properties of Soil Using Artificial Neural Network

Abstract

This study focuses on expediting the determination of the soaked California bearing ratio (CBR) in laboratory assessments, which traditionally demand substantial time and effort. This research endeavors to create predictive models for CBR in Addis Ababa city using artificial neural network modeling techniques. These models are developed based on input parameters encompassing Gravel percentage (G), Sand percentage (S), Fines percentage (F), Liquid limit (LL), Plastic limit (PL), Plasticity Index (PI), maximum dry density (MDD), and optimum moisture content (OMC).The study begins with a sensitivity analysis to identify influential parameters, as there exist varying perspectives among researchers regarding which parameters carry the most significance for predicting the soaked CBR of fine-grained soils. Initial simple regression analyses conducted in this study reveal that Fine percentage and PI exhibit stronger correlations with soaked CBR of fine-grained soils, with correlation coefficients of 0.66 and 0.47, respectively, compared to maximum dry density and optimum moisture content. Based on these correlations, neural network and multiple regression prediction models are developed, incorporating varying numbers of input parameters. The results demonstrate that neural network models excel in utilizing less influential parameters, in contrast to multiple regression models. Neural network analysis proves particularly advantageous in situations where uncertainty exists regarding the most influential input parameters, as it enables the simultaneous consideration of all such influential parameters, thereby enhancing model performance. The research employs a multi-layer perceptron network with feed-forward back propagation, utilizing a dataset comprising 318 primary and secondary soil test results for index properties and CBR, conducted in accordance with ASTM Standards. During testing, the developed models exhibit noteworthy success in predicting CBR, achieving correlation values of approximately 0.9002. The models undergo cross-validation using primary soil test data. Ultimately, the study's findings underscore the superiority of the artificial neural network approach in predicting CBR, outperforming multiple regressions, which yielded a correlation coefficient of 0.75 for the combination of Gravel, Fine, Sand, and PI.