Prediction of Target LWD Modulus of Unbound Geomaterials by Lab LWD Test

Abstract

Modulus-based compaction acceptance evaluation of geomaterials using lightweight deflectometers (LWDs) had gained increasing attention in recent decades, as the traditional density-based method using nuclear density gauges (NDGs) was related to safety, regulatory, and cost concerns. Missouri Department of Transportation was interested in adopting LWDs in place of NDGs and creating related standards. In this method, determination of target LWD modulus was needed, because compaction acceptance was evaluated by comparing the in-situ LWD modulus with a predetermined target one. This study aimed to apply lab LWD test (LLWDT) method to characterize the relationship between lab LWD modulus, moisture content (MC), and applied stress for target LWD modulus prediction. With data obtained from a matrix of LLWDTs conducted using various soils, MCs, and applied stresses, the characterization was first performed for each soil using the model given in the American Society for Testing and Materials (ASTM) standard. However, this model illustrated a parabolic shape and failed to capture the inverse-sigmoid pattern shown in the data. Given the identified limitation, two machine learning (ML) models including random forest and multilayer perceptron (MLP) were employed for the characterization and compared with the ASTM model. The comparison suggested that the two ML models exhibited comparable prediction performance and considerably outperformed the ASTM model, with MLP showing the best generalization capacity.