Temperature-Dependent Transport and Mechanical Properties of Frozen Soils: Physics-Based Parameterization and Data-Driven Modeling

Abstract

Frozen soils behavior is governed by strongly coupled thermal, hydraulic, and mechanical processes driven by temperature-dependent phase changes and unfrozen water redistribution, which collectively control transport properties, mechanical response, and the performance of infrastructure in cold regions. Despite extensive experimental and theoretical studies, reliable determination of thermo-hydro-mechanical (THM) properties remains challenging, as experimental investigations are time-consuming, require specialized facilities, and are often prohibitively expensive. Recent advances in machine learning (ML) offer a complementary and cost-effective strategy for parameterizing frozen soil properties by utilizing existing experimental data reported in the literature. The primary objective of this thesis is to develop an integrated and scalable modeling framework for frozen soils by combining physics-based formulations with data-driven methods to predict temperature-dependent transport and mechanical properties and their coupled behavior.

The first goal of this thesis is to establish a transferable predictive framework for temperature-dependent transport properties that include thermal conductivity and hydraulic conductivity. To this end, physics-informed and data-driven models are developed to capture the nonlinear dependence of transport properties on temperature, unfrozen water distribution, and pore-scale liquid connectivity across a wide range of soil textures. The results demonstrate that transport behavior in frozen soils is jointly governed by phase-change-controlled liquid distribution and pore-scale connectivity, indicating that neither thermal conductivity nor hydraulic conductivity can be uniquely characterized by unfrozen water content alone. The soil texture and pore-scale ice-water configuration play a critical role in controlling heat transfer and liquid-phase connectivity and help explain to a greater extent the observed variability of traditional models. In addition, the evolution of liquid connectivity during freezing, together with freezing pattern within pores, governs the temperature sensitivity of hydraulic conductivity, leading to markedly different hydraulic transport capacities under identical unfrozen water contents.

In parallel, the second goal of this thesis is to develop a high-dimensional and generalizable data-driven framework for modeling the mechanical response of frozen soils. A hybrid modeling strategy is proposed to describe the highly nonlinear stress-strain behavior of frozen sands under varying temperature, confining pressure, and loading conditions. The results reveal that although frozen soil behavior exhibits high-dimensional characteristics, its mechanical response is strongly influenced by a limited number of key state variables, while remaining inherently path- and temperature-dependent. These findings highlight the necessity of modeling approaches that explicitly account for loading history and uncertainty to achieve robust and transferable predictions.

The third and final goal of this thesis is to integrate the developed transport and mechanical models into a unified THM numerical framework for engineering-scale applications. The coupled framework enables simulation of freezing-induced processes, including freezing front evolution, ice accumulation, and soil deformation under transient thermal and hydraulic conditions. Coupled THM simulations further demonstrate that consistent parameterization of temperature-dependent transport and mechanical properties is essential for capturing the coupled evolution of thermal, hydraulic, and mechanical responses in frozen soils.

In summary, this study demonstrates that accurate prediction of frozen soil behavior requires a modeling framework that explicitly represents pore-scale freezing geometry and liquid-phase connectivity, rather than relying on unfrozen water content as a sole state variable. The proposed framework provides a robust basis for coupled THM modeling of frozen soils, offering practical potential for improving the design and analysis of cold-region infrastructure while reducing reliance on time-consuming and costly experimental testing.