ARTIFICIAL NEURAL NETWORKS FOR ESTIMATION OF SOIL-WATER CHARECTERISTIC CURVE

Abstract

Soil-Water Characteristic Curve (SWCC) is one of the fundamental aspects in unsaturated soil mechanics which describe hydraulic conductivity and soil behavior at different levels of moisture. This study aims to create Artificial Neural Network (ANN) deterministic model to predict SWCC fitting parameters according to Fredlund and Xing (1994) equation. A combination of external databases and experimentally obtained data points of 10 soil samples were utilized to train and validate model performance. This model established a connection between input parameters such as grain-size distribution (GSD) and Atterberg limits to predict output parameters such as a, n, m and 𝜃s values. The results of the research demonstrated high level of accuracy in prediction of SWCC curve, where performance matrices such as RMSE and coefficient of determination (R2) showed 0.023 - 0.0863 and 0.869 – 0. 987 respectively. In comparison with traditional methods of SWCC estimation, ANN showed superior prediction capabilities by excluding time-consuming and expensive steps. However, current research highlighted the need of extending data bank for future models to achieve better prediction results for all soil types including sandy, silty and clayey soils. Generally, this research contributes to the advancement of AI approach in geotechnical engineering, which potentially could offer efficient and accurate alternative to existing empirical methods.