Predicting the Unfrozen Water Content of Freezing Soils Using an Artificial Neural Network Model
Jun Bi, Yuxuan Pan, Wenxuan Mu, Sheng Qiang Yang, Guoxu Wang, Mengyao Mao, Shengnian Wang, Tingting Wei
Abstract
Unfrozen water content is one of the most prominent hydrothermal properties because it largely affects the hydraulic, physical, and mechanical behavior of frozen soils. However, the accurate estimation of unfrozen water contents under different subzero temperatures is challenging because it is affected by various soil properties. In this study, an artificial neural network (ANN) model (feedforward neural network) was proposed to estimate the unfrozen water content. A database with 1,033 unfrozen water content measurements and their corresponding influencing soil parameters was compiled from 16 published articles. The influencing soil parameters were ranked based on the Spearman correlation coefficient. After that, we investigated the effects of input soil properties and the numbers of neurons in the hidden layer on the estimated results. The unfrozen water contents estimated by the ANN models were evaluated with experimental results and four empirical models. Results suggested that the developed ANN models outperformed the empirical models, indicating that the ANN model has a superior ability to estimate the unfrozen water content of frozen soils.