Soft Computing-Based Prediction of CBR Values

California Bearing Ratio method is an empirical method of design of flexible pavement developed by California Division of Highways, in 1928 for the design of Roadways, Railways and Airfield. In order to design a pavement by CBR method, the soaked CBR value of soil is evaluated which takes around 4 days or 96 h to complete the test process. The soaked CBR value is used to determine the total thickness of flexible pavement needed to cover the subgrade of the known CBR value. However, the determination of soaked CBR in the laboratory is time-consuming and requires skilled labour and supervision, prompting researchers to explore alternating approaches. Various machine learning methods including artificial neural network (ANN), deep neural networks (DNN) and gene expression programming (GEP) have been previously employed to predict CBR values. However, these methods come with inherent limitations such as sensitivity to hyper-parameters, limited flexibility, lack of interpretability and explainability which raise concerns in critical decision-making applications. In the present study, we have endeavoured to address the shortcomings observed in deep neural networks models and proposed an improved and efficient prediction model for California Bearing Ratio. Three distinct models have been developed using three different methodologies: a fuzzy inference system, an artificial neural network & an adaptive neuro-fuzzy inference system. To conduct the study, large datasets of 2000 Soil samples have been used which were tested under the scheme of Pradhan Mantri Gram Sadak Yojana (PMGSY). Out of the total data, 1501 datasets have been used for training, 499 datasets have been used for testing and for validation of the proposed model, datasets of 15nos soil samples have been used which were entirely separated from the datasets used for training and testing. Upon analysing the prediction results, we found that while ANN demonstrates commendable accuracy in predicting CBR values, the predictability of the manually developed FIS model falls short. Intriguingly, the ANFIS model surpasses both ANN and FIS in terms of predictive accuracy with an a-20 index of 0.83 and R values of 0.92. In Conclusion, our study suggests that the hybrid model of ANN and FIS (ANFIS) emerges as a promising approach for predicting CBR values, offering enhanced accuracy compared to traditional methods and other machine learning models.