Statistical analysis and prediction of spatial resilient modulus of coarse-grained soils for pavement subbase and base layers using MLR, ANN and Ensemble techniques

Abstract

Prediction methods for the estimation of the resilient modulus({M}_{R}) of unbound granular materials continues to be fraught with challenges associated with the in situ dependency of the existing methods, which makes them unreliable when considering the spatial variability of soil properties. Moreover, the artificial intelligence methods, which prove to be superior for {M}_{R} prediction, indicate that more recent and hybrid machine learning methods are more apt. Consequently, a comparative analysis was executed in the present study using advanced statistical approaches and a simple routine method for the prediction of the {M}_{R} of coarse-grained soils used as unbound granular materials in highway pavements. This study considered also the spatial variability of the routine soil properties based on the long-term pavement performance database. The result of the descriptive analysis conducted using principal component analysis revealed that the variation in {M}_{R} is due to three principal factors, which include the effect of moisture on the fines content, effect of coarse particles on the compaction characteristics and the effect of the soil stress state. Subsequently, prediction of the {M}_{R} using multiple linear regression was unreliable based on residual analysis and multi-collinearity problems. Thus, predictive modeling of {M}_{R} was then executed using artificial neural network (ANN) and ensemble techniques (random forest and gradient boosting machine), which resulted in very high R2 values above 0.9, considered apt for MEPD. The ANN model was found to provide a superior prediction performance than the ensemble techniques.