Resilient modulus descriptive analysis and estimation for fine-grained soils using multivariate and machine learning methods

Abstract

ABSTRACT The adoption of mechanistic-empirical approach to pavement design requires the use of resilient modulus of subgrade soils as a crucial input. The determination of in the laboratory is inexpedient due to the nature of the existing test protocols. This prompted the use of estimated values, which inadvertently has gained popularity lately. However, the accuracy of estimated values is questionable due to spatial variability of soil properties. This necessitated the aggressive search for robust and thorough approaches for predictive modelling of the . In the present study, a systematic approach was adopted for the descriptive analysis and estimation of . from routine soil properties using data from Long-Term Pavement Performance (LTPP) and considering the spatial variability of the soil properties. Descriptive analysis was executed using non-parametric correlation and principal component analysis (PCA), while the estimation was done using three machine learning methods which include gradient boosting regression (GBR), adaptive neuro-fuzzy inference system (ANFIS) and artificial neural network (ANN). Based on the PCA, four factors which explained a total of 77.5% variance in the data had significant influence on the . These include the effect of moisture-induced changes on the soil consistency limits and physical condition, effect of the soil clay content, effect of the soil gradation and effect of the soil stress state. Various factors of the machine learning methods such as the learning rate, number of clusters and number of hidden layers had a significant effect on the prediction accuracy. The three machine learning methods were satisfactory for the prediction based on R 2 values which were generally above 0.9. Also, when considering spatial variability of routine soil properties, the GBR and ANFIS have a comparative advantage over the ANN, since they exhibited a high stability in the prediction for both the training and testing dataset.