A deep neural network model to estimate water retention of compacted clayey soils

Soil-water retention is fundamental to understand hydro-mechanical characteristics of unsaturated clayey soils. The soil-water retention curve (SWRC) depends on internal (e.g. mineralogical composition, and chemo-physical properties of soils) and external (e.g. stress states and temperature) factors. The SWRC is usually determined through laboratory testing, which is costly and time consuming. In this study, we compiled an experimental dataset containing water retention data of artificial and natural clayey soils to develop a deep neural network (DNN) model trained with genetic algorithm (GA) to estimate SWRC over a wide suction range. The relevant soil properties including dry density, liquid limit, plastic limit, plasticity index, initial water content, void ratio, and suction are the input variables of the DNN-GA model, while the gravimetric water content is the output variable. The analysis of modeling errors and the comparison of gravimetric water content predicted values with experimental values showed the high efficiency of the model being developed. The DNN-GA model can be used as an accurate alternative to classical soil mechanic correlations.