Validation of Compression Index Approximations Using Soil Liquid Limit

Abstract

The list of published empirical models for approximating the compression index of fine-grained soils in terms of its liquid limit continues to grow. Many such expressions are justified based on high correlation coefficients using limited test data sets. Significant measures such as standard errors are often omitted. Some published models are applicable to fine-grained soils, organic soils or all soils. Most of the published models are based on traditional linear regression which ignores data outliers. Artificial Neural Networks (ANN) regression are being used to develop more reliable models but often ignores data outliers. Although ANN is a valuable tool, it is data driven and the quality and quantity of the data impact developed models. Over the past five years, significant data was collected from published articles from reputable journals and conferences. A total of 1906 data points relating soil compression index to its liquid limit were analyzed using Robust Bi-Square regression to reduce the impact of data outliers. Published models rely on traditional regression analysis. The results indicate several published data sets involving liquid limits of less than 16 while others included data that plotted above the U-line in the Plasticity Chart. Such data points are impossible to exist in engineering practice. Using MATLAB and the Robust Bi-Square regression, the authors developed new regression models with high correlations and low standard errors that relate the compression index of fine-grained soils to its liquid limit. The proposed models are more reliable than those developed using traditional regression technique which ignore data outliers and often involve incorrect data. The proposed Robust regression models were compared with published empirical approximations that are based on traditional regression methods to establish validity and reliability.