Predictive models for mechanical properties of expanded polystyrene (EPS) geofoam using regression analysis and artificial neural networks

Abstract

Initial elastic modulus and compressive strength are the two most important engineering properties for modeling and design of EPS geofoams, which are extensively used in civil engineering applications such as light-fill material embankments, retaining structures, and slope stabilization. Estimating these properties based on geometric and physical parameters is of great importance. In this study, the compressive strength and modulus of elasticity values are obtained by performing 356 unconfined compression tests on EPS geofoam samples with different shapes (cubic or disc), dimensions, loading rates, and density values. Using these test results, the mechanical properties of the specimens are predicted by linear regression and artificial neural network (ANN) methods. Both methods predicted the initial modulus of elasticity (E-i), 1% strain (sigma(1)), 5% strain (sigma(5)), and 10% strain (sigma(10)) strength values on a satisfactory level with a coefficient of correlation (R-2) values of greater than 0.901. The only exception was in prediction of sigma(1) and E-i in disc-shaped samples by linear regression method where the R-2 value was around 0.558. The results obtained from linear regression and ANN approaches show that ANN slightly outperform linear regression prediction for E-i and sigma(1) properties. The outcomes of the two methods are also compared with results of relevant studies, and it is observed that the calculated values are consistent with the results from the literature.