Data-driven ANN model for estimating unfrozen water content in the thermo-hydraulic simulation of frozen soils

Abstract

This study integrates a data-driven model for estimating the unfrozen water content into the thermo-hydraulic coupling simulation of frozen soils. An artificial neural network (ANN) was employed to develop this data-driven model using a dataset from the literature. Thereafter, a numerical algorithm was developed to implement the data-driven model into the thermo-hydraulic simulation. In the numerical algorithm, the frozen and unfrozen zones are distinguished first according to the freezing temperature, where the unfrozen water at frozen nodes is updated using the ANN model. Subsequently, discretized hydraulic and thermal equations are solved sequentially and iteratively using Newton-Raphson method until the temperature and unfrozen water content satisfy the tolerance simultaneously. Horizontal and vertical freezing experiments are used to verify the reliability of the proposed algorithm. The computed variations in temperature, total water, unfrozen water, and ice content achieve good agreements with measured data. Some key features of frozen soils, such as water migration and ice formation, and the increase in total water content, are reproduced by the developed algorithm. Additionally, the comparison between the ANN model and existing empirical equations for determining unfrozen water content demonstrates that the ANN model offers a better performance.