Soft Computing Models to Predict the Compaction Characteristics from Physical Soil Properties

Abstract

In almost every earthwork, it is essential to compact soil so that the densest possible state of the soil can be achieved. The suitability of soil for earthworks relies on the compaction characteristics; Optimum Moisture Content (OMC), and Maximum Dry Density (MDD). The determination of the compaction characteristics in the laboratory, for a vast volume of soil, is time-consuming. Therefore, for the initial assessment of soil, it is crucial to determine the compaction characteristics from physical soil properties. In this work, three different models of the Artificial Neural Network (ANN), M5P-tree, and Multiple Linear Regression (MLR) are used to predict the compaction characteristics of the soil. In the models, particle size and plasticity properties of soil are combined and, seven input parameters of gravel, sand, silt, and clay contents, plastic limit, liquid limit, and plasticity index are comprised. To develop the models, 1038 datasets are compiled and processed. Several statistical analyses, including coefficient of determination (R2), scatter index (SI), root mean squared error (RMSE), mean absolute error (MAE), and Objective (OBJ) value, are used to assess the effectiveness of the proposed models. The findings demonstrated that overall, the ANN model performed better in predicting the OMC, while the MLR model performed better in predicting MDD. Further, from the sensitivity analyses, it was indicated that the plastic limit has more influence on the value of OMC while, to predict the MDD, both sand content and the plasticity index play a foremost role.