AI-Based Estimation of Swelling Stress for Soils in South Africa

Abstract

Soil swelling is one of the major causes of structural failure, due to excessive moisture saturation and desaturation moisture. In situ measurement of swelling stress is somewhat impossible and requires tedious routine site observation. The use of artificial intelligence to predict the swelling stress of in situ soil is highly recommended, because of the complex behavior of soil upon moisture absorption. Because of this challenge, this study is channeled towards the prediction of swelling stress using basic geotechnical properties to save the challenges of repeating geotechnical experimental tests. In this study, the swelling stress of soils collected from 15 locations in 5 sites across South Africa has been predicted by using the artificial neural network (ANN), genetic programming (GP), and evolutionary polynomial regression (EPR)-based intelligent techniques. Multiple data were collected through laboratory experiments on the predictors: gravimetric moisture content (GMC), plasticity index (Ip), dry density (γd), free swell index (FSI), degree of saturation (S), matric suction (Ψm) and the target, and swelling stress (Psm). This predictive model was aimed at proposing models, which will help earthwork designers and constructors in South Africa overcome the rampant visit to the laboratory in search of soil data needed for geotechnical engineering designs. The soils showed their potential for swelling, which was eventually confirmed by the sensitivity analysis of the intelligent models. The performance indices of the models showed that ANN outclassed the other techniques with a performance accuracy of 93.6% at an error of 1.9%. Also, the sensitivity analysis showed that the plasticity index and matric suction were the most influential to the models. With the predicted models, future earthworks in South Africa can quickly forecast swelling stress prior to designs and construction more so in hydraulically bound environments.