Assessing the swelling/shrinkage potential of Italian soils via artificial neural network approach

For the prediction of the swelling/shrinking parameters (swelling pressure and swell strain) a multi-layered feed-forward neural network, trained using an error-backpropagation algorithm, is developed.Two levels of input variables were used: the simpler models consider only parameters that depend on the nature of the soil; the complex models take into account also variables that are influenced by environmental factors and climatic conditions. The data for this study was taken from a laboratory-based database. The data was grouped in relation with the genesis of the soils. Results indicate that complex neural network models provide significant improvement for prediction of soil swelling/shrinking parameters over those predicted via the simpler ones. Additionally, it was noted that the developed models performed well when addressing data sets concerning one geologically based soil. For this reason, it is recommended that individualized models need to be developed for each geologically based soil-type.