Thermal conductivity is one of the important properties required for understanding the frozen soils behavior. There are several models available in the literature for the prediction of thermal conductivity of frozen soils based on the proportions of unfrozen water, ice, gas, and soil particles. In this study, two ensemble learning methods-based models; namely, the Random Forest (RF) model and the Least Squares Boosting (LSB) model, are extended to estimate the thermal conductivity of frozen soils. These models utilize basic soil properties as input parameters that include water content, dry density, temperature, and fractions of gravel, sand, silt, and clay, can be measured easily, or determined. Additionally, seven widely used thermal conductivity models, referred to as the traditional models for frozen soils, were evaluated. Both the RF and LSB models, as well as the traditional models, were assessed using data of 823 tests derived from 43 soils with different textures that were gathered from the literature. The results highlight that the traditional models have their strengths and limitations in terms of their use for different types of soils. In contrast, the proposed ensemble learning methods-based models provide higher prediction accuracy compared to the traditional models and can be applied to all soil types and temperature ranges. Furthermore, estimation from the ensemble learning methods-based models can be used to provide probability of multi-dimensional analysis of frozen soils.
Read full abstract