Neural network based model for estimation of the level of anisotropy of unbound aggregate systems

Directional dependency of the materials properties in granular soils is an integral component of the analysis and design of pavement foundations. Laboratory determination of such anisotropic properties is often overlooked by design engineers due to the lack of available protocols, equipment, time constraints, and budgetary issues. This research strives to bridge this gap by investigating an alternative approach to provide a practical and reliable framework to estimate the level of anisotropy of unbound granular materials. To achieve this objective, an experiment design was developed to establish a comprehensive aggregate feature database. Nonlinear and cross anisotropic material properties were determined using Variable Dynamic Confining Pressure (VDCP) stress path tests in the laboratory. Particle geometry was characterized using the Aggregate Imaging System (AIMS). Scale parameters and shape parameters of the form, angularity and textural properties of the particles were incorporated in the aggregate database to account for the shape-induced anisotropy of particulate soils. Moisture state and density parameters were also incorporated in the database for further post processing. Several neural network models with different architectures were developed, and the performances of the models were assessed based on an unseen set of data. The impact of the neural network topologies on the performance of the models were also investigated in this research effort. The results indicate that increasing the number of hidden layers has negligible impact on the performance of the models. Additionally, a series of sensitivity analyses were performed to investigate the level of contribution of aggregate features on the anisotropic behavior of unbound granular materials. The results of the sensitivity analysis underscores the influence of aggregate angularity, fines content, surface macro-texture, as well as the parameters of the stress path tests, on the orthogonal load distribution capacity of aggregate layers. (C) 2018 Elsevier Ltd. All rights reserved.