Prediction of Stress-Dependent Soil Water Retention Using Machine Learning

The soil water retention curve (SWRC) provides information for a wide range of geoenvironmental problems, such as analyses of transient two-phase flow, the bearing capacity and shear strength of unsaturated soils. Many past studies have shown experimentally the effects of stress on the SWRC. Unfortunately, direct stress-dependent water retention measurements are relatively time-consuming and generally require special equipment and a certain level of expertise. This study primarily aimed to develop a novel predictive framework within the context of soft computing to capture the dependency of the SWRC on several variables, with an emphasis on stress and soil type. To achieve this, the three shape parameters of van Genuchten's water retention model were estimated using a comprehensive database of 102 SWRC tests retrieved from the literature. In this study, 60% of the datasets were employed for model training, with an additional 20% being designated for validation, while the remaining 20% were set aside for testing the model's performance. The data were analyzed using two machine learning techniques: the group method of data handling and multi-layer perceptron approaches. Results showed excellent performance of the two methods. A sensitivity analysis was conducted to explore the relative significance of the different variables. Interestingly, net stress was found to be almost as significant as soil type. The introduced artificial intelligence based predictive framework provided a very effective method of integrating theory and practice.