Prediction of the resilient modulus of non-cohesive subgrade soils and unbound subbase materials using a hybrid support vector machine method and colliding bodies optimization algorithm

Abstract

The resilient modulus (MR) of road materials is one of the most important parameters in the analysis and design of pavement. This parameter is used in both empirical methods and mechanistic-empirical methods as the main parameter for expressing the stiffness and behavior of road construction materials. To determine this parameter in the laboratory, it is necessary to perform a dynamic tri-axial loading test under various confining and deviator stresses, which is a time- and cost-intensive approach. In this paper, a support vector machine (SVM) hybridized with the colliding bodies optimization (CBO) algorithm was used to model the MR of non-cohesive subgrade soils, unbound subbase materials, and subgrade-subbase materials. For this purpose, a dataset was collected from published literature, which included 190 distinctive data records on 13 different types of subbase materials and 272 distinctive data records on 18 different types of non-cohesive subgrade materials. To develop and validate the predictive models for MR, 75% of the data records were selected as training dataset while the remaining 25% were selected as testing dataset. The input variables in all the models included maximum dry density, uniformity coefficient, curvature coefficient, percent passing No. 200 sieve, confining stress, and deviator stress. For the models developed for subbase, subgrade, and subbase-subgrade materials, the R2 values obtained for the testing dataset were 0.988, 0.979, and 0.972, respectively. The results of this study also indicate that the SVM-CBO method is more accurate than the ANN method in predicting the MR of non-cohesive subgrade soils and unbound subbase materials. Moreover, a series of gamma tests were conducted to evaluate the most effective parameters for the prediction of the MR. For the subbase-subgrade materials, percent passing No. 200 sieve and confining stress were found to be the least and the most effective variables. A parametric analysis was also done to study the effects of each input variable on the MR.