This study investigates the thermal conductivity (λ) and volumetric heat capacity (C) of sandy soil samples under a variety of conditions, including freeze-thaw cycles at temperatures both above and below zero and differing moisture levels. To estimate these thermal properties, a novel predictive model, EFAttNet, was developed, which utilizes custom-designed embedding and attention-based fusion networks. When compared to traditional de Vries empirical models and other baseline algorithms, EFAttNet demonstrated superior accuracy. Preliminary measurements showed that λ values increased linearly with moisture content but decreased with temperature, whereas C values exhibited a rising trend with both moisture content and freezing temperature. Following freeze-thaw cycles, both λ and C were positively influenced by moisture content and freezing temperature. The EFAttNet-based model proved highly accurate in predicting thermal properties, particularly effective at capturing nonlinear relationships among the influencing factors. Among these factors, the degree of saturation had the most significant impact, followed by the number of freeze-thaw cycles, subzero temperatures, porosity, and moisture content. Notably, dry density exerted minimal influence on thermal properties, likely due to the overriding effects of other factors or specific soil characteristics, such as particle size distribution or mineralogical composition. These findings have significant implications for construction and engineering projects, especially in terms of sustainability and energy efficiency. The demonstrated accuracy of the EFAttNet-based model in estimating thermal properties under various conditions holds promise for practical applications. Although focused on specific soil types and conditions, the insights gained can guide further research and development in managing soil thermal properties across diverse environments, thereby enhancing our understanding and application in this field.
