Prediction of Engineering Properties of Soils from Index Properties using Artificial Neural Network

Abstract

The major difficulties associated with the determination of engineering properties of soil are the difficulty in obtaining undisturbed samples and most of the methods are rather expensive, time consuming and involve elaborate test procedures. On the other hand, tests to determine index properties are quite simple and less expensive. Hence, there have been many attempts to correlate or predict the engineering properties from the index properties by many researchers. These methods are usually purely empirical or semi empirical which are generally based on past experience, field observations etc. And the reliability or range of accuracies of these predictions is widely varying. Thus, it would be ideal to evolve some other method to predict the engineering property of soil from the basic index properties with a higher degree of accuracy or lesser value of error. The ANN is an effective technique that can be used to solve complex problems involving many factors with different degrees of accuracies or confidence levels. In this present work, a method for predicting engineering properties of soil from index properties using artificial neural network is developed. The database of 500 soil sample details used for the development of the ANN were based on the various soil investigation works and laboratory tests conducted by the Geotechnical Engineering division of the Civil Engineering department of NIT Calicut. The major input parameters are fine fraction, liquid limit, plasticity index, natural density, natural water content, D₁₀. The output parameters like shear strength parameters (c, ϕ), compressibility parameter (Cc) and permeability are computed by the network. ANN predicted outputs were then compared with traditional methods for predicting angle of friction, cohesion, compression index, and permeability. It was observed that ANN model developed could predict the engineering properties better than other traditional methods.