Artificial Neural Network model to predict the dynamic properties of sand-polyurethane composite materials for GSI applications

Abstract

This study shows the modelling of the dynamic properties of sand-polyurethane composite materials through the Artificial Neural Network method (ANN). Such properties are relevant to assess the effectiveness of polyurethane injected into the foundation soil as Geotechnical Seismic Isolation (GSI) system. Three FeedForward Neural Network (FFN) models are developed for the prediction of the shear modulus decay, the normalised shear modulus decay and the damping ratio curve. Four inputs are considered, namely the dynamic properties of pure sand and pure polyurethane, the volumetric percentage of polyurethane and the isotropic confining stress. Each model is trained and tested on a database including previous experimental results of Resonant Column (RC) tests performed on layered sand-polyurethane specimens, with polyurethane volumetric percentages varying from 10% to 30%. In all cases, the network architecture is found to be optimal with a single hidden layer and the number of hidden neurons from 4 to 6 according to the complexity of the data to be fitted. By comparing the predicted properties with the experimental results, a good prediction quality of the models is shown both in terms of the coefficient of correlation R2 and the Mean Squared Error MSE. A relative importance analysis finally reveals that polyurethane mainly affects the properties of the composite materials with its volumetric percentage, rather than its dynamic properties especially in terms of shear modulus.