Modeling time dependent swell of clays using sequential artificial neural networks

This work attempts to implement sequential artificial neural networks (SANN) for modeling time dependent swell of expansive soils. Forty soils with varying properties were selected and tested for expansion under three different initial applied pressures (25, 100, and 200 kPa) to develop the database used for training and testing the neural network. Consequently, a total of 120 swell tests were performed to produce over 1800 data points. The input parameters used in the network included the soil initial dry unit weight and water content, initial applied pressure, percent clay content, plasticity index and the percent swell at time i. The network was programmed to process this information and produce the percent swell at time i + 1. The study demonstrates that there is a possibility to develop a general SANN model that can predict time dependent swell, based on basic soil properties, with relatively high accuracy (correlation r(2) = 0.975 between predicted and observed data).