Prediction of UCS and CBR of microsilica-lime stabilized sulfate silty sand using ANN and EPR models; application to the deep soil mixing

Abstract

Desert sands in Iran, which usually contain small amounts of silt and sulfate, do not have significant strength, and thus, are not suitable for foundations or road construction. This paper applies the results of 90 Unconfined Compressive Strength (UCS) and California Bearing Ratio (CBR) tests on sulfate silty sand stabilized with different lime and microsilica percentages as the two main stabilizers. Based on the obtained databank from the tests, Back Propagation Artificial Neural Network (BP-ANN) and Evolutionary Polynomial Regression (EPR) models are developed to predict the UCS and CBR values. Assessing the different architectures (one- and two-hidden layer neural networks) and functions (polynomial, exponential and hyperbolic tangent functions for the EPR models), a BP-ANN model with 5-5-8-1 layers and an EPR model with a hyperbolic tangent function showing high accuracy are introduced as the best models for predicting the UCS. Through a sensitivity analysis, the most and the least influential parameters on the UCS are presented and the results are further discussed using scanning electron microscopy (SEM). The presented EPR models can be useful for practitioners when selecting the optimized percentage of stabilizers or for controlling purposes in the QC/QA phases of deep soil mixing projects. In this regard, the application of the proposed models to the design of deep soil mixing is presented and elaborated using an example. In this example, the optimum and the best practical amounts of stabilizers are obtained through the graphical optimization of the models. In addition, by applying the developed relationships to a new case, the comprehensiveness of the developed relationships is further declared and it is shown that the proposed relationships are practical and can be efficiently used in the preliminary design stage.