Investigation of the thermal conductivity of soil subjected to freeze-thaw cycles using the artificial neural network model

Abstract

In cold regions, a better understanding of soil thermal conductivity is necessary for a variety of earthworks and engineering applications such as ground heat exchanger piles and energy piles, installation of underground power cables, and so on. In this study, the effects of the internal factors of the soil such as water content, dry density, porosity, saturation degree, and the external factors of the soil like freeze-thaw cycles and temperatures were studied on the thermal conductivity (lambda) of the sandy soil. The lambda values of the soil samples were determined at six different volumetric water contents (0.190, 0.212, 0.230, 0.246, 0.260, and 0.320 m(3) m(-3)) and four frozen temperatures (4 degrees C, - 7 degrees C, - 12 degrees C and - 20 degrees C) under five different numbers of freeze-thaw cycles (0, 2, 5, 9, and 12). Then, new prediction models based on the internal (the ANN-I model) and both internal and external factors (the ANN-G model) of the soil proposed by artificial neural network (ANN) technology. The developed ANN models were compared with three empirical models (Tortuosity-Parallel model, Farouki model, and de Vries model) to verify their reliability and effectiveness. The results showed that dry density, water content, porosity, saturation degree, and temperature have significant and variable influences on the lambda of the soil subjected to repeated freeze-thaw cycles. The ANN-G model provided the highest accuracy in predicting the lambda with increasing numbers of freeze-thaw cycles. For frozen sandy soil samples subjected to repeated freeze-thaw cycles, the Tortuosity-Parallel model exhibited the best performance, followed by the Farouki model and the de Vries model with the poorest performance.