Wang, Junjie2025-11-262025-11-262025-11-26http://hdl.handle.net/10393/51102https://doi.org/10.20381/ruor-31558Compacted unsaturated soils are widely used in the construction of embankments, pavements, subgrades, foundations, retaining walls, railways and slopes, and their hydro-mechanical behavior is influenced by both net normal stress and matric suction. Although the theoretical framework of unsaturated soil mechanics is well established, routine determination of matric suction, the soil water characteristic curve (SWCC), and related hydro-mechanical properties remain labor-intensive, slow and often impractical for engineering practice applications. This difficulty is more pronounced in compacted soils because microstructural effects and slow suction equilibration complicate both measurement and interpretation. These challenges motivated the development of simplified methods by various researchers for predicting or estimating hydro-mechanical properties in many regions over the last three decades. In these methods, SWCC, which is defined as a relationship between the water content (i.e., volumetric or gravimetric) or degree of saturation and soil suction, has been used as a tool along with saturated soil properties for reliably predicting the variation of unsaturated soil properties with respect to suction. This background is widely used in the design and modeling of the behavior of various geo-infrastructure. Nevertheless, studies related to assessing the in-situ behavior of various infrastructures remain a significant limitation. For example, reliable and continuous measurement of in-situ matric suction has been a challenge. This limitation also continues to hinder the in-situ measurement of SWCC, which is useful to assess the performance of geo-infrastructure. Another example is natural precipitation that can alter matric suction through water infiltration, leading to variations in unsaturated soil properties. Reliable information of matric suction, along with the other hydro-mechanical properties is required for the reliable hazard assessment of geo-infrastructures. Such information can be useful in suggesting correction and maintenance/repair measures to alleviate possible geotechnical failures. To address some of these challenges, a critical literature review is conducted within the scope of the research, focusing on the measurement techniques and prediction methods for matric suction and the SWCC. Building on insights from this review, this thesis proposes a series of advanced machine learning (ML) frameworks that integrate data-driven modeling with the fundamental principles of unsaturated soil mechanics. The research addresses four interrelated challenges in the estimation and prediction of key properties of unsaturated soils. The first objective aims to estimate matric suction from soil properties that are routinely obtained through standard laboratory testing. Two prediction models are developed: namely, Particle Swarm Optimization Support Vector Regression (PSO-SVR) and Multivariate Adaptive Regression Splines (MARS) are developed in this study. PSO-SVR demonstrates that the soil matric suction can be predicted with a reasonably high accuracy, while MARS facilitates efficient input selection and sensitivity analysis. By combining their strengths, an integrated framework is proposed and validated using published datasets of matric suction from literature, particularly for low-plasticity soils within a suction range of 0 to 1500 kPa. This approach reduces reliance on direct matric suction measurement and supports more rapid risk assessments of geotechnical infrastructure. The second objective addresses the specific challenges associated with fine-grained compacted soils, whose hydro-mechanical behavior is highly sensitive to microstructural characteristics. In practice, measuring matric suction in these soils is considerably more complex and time-consuming than in low-plasticity soils, due to slower equilibration rates, greater structural sensitivity, and the need for specialized high-capacity devices. Recognizing these difficulties, this study develops a hybrid ML framework that integrates PSO-SVR with Multi-Gene Genetic Programming (MGGP). This approach leverages the high predictive capability of PSO-SVR and the transparent, interpretable structure of MGGP, enabling both robust estimation and empirical formula derivation. A novel ML-based parameter, termed the effective degree of aggregation, is introduced to quantitatively represent the influence of soil structure and aggregation under varying initial water contents, thereby improving the model's responsiveness to microstructural variations. Sensitivity analyses are conducted to identify the most influential input features, ensuring the resulting models are both physically meaningful and practically applicable. The explicit empirical equations derived from the MGGP component facilitate direct use in engineering applications without requiring complex computational tools. The framework is validated against measured data, showing strong agreement and demonstrating its utility for the practical estimation of matric suction in compacted fine-grained soils. The third objective focuses on predicting the SWCC, a widely adopted and efficient tool in geotechnical design for estimating hydro-mechanical properties of unsaturated soils. Traditional methods for determining the SWCC often rely on extensive laboratory testing and curve-fitting procedures, which can be time-consuming and sensitive to data variability, particularly in compacted fine-grained soils. To address these limitations and improve modeling robustness, a hybrid ML framework is developed by integrating Extreme Gradient Boosting (XGBoost) and Multi-Layer Perceptron (MLP) algorithms with thermodynamic principles. This framework approximates the Fredlund and Xing (1994) (FX) SWCC model by relating its fitting parameters to fundamental soil properties through ML regression. A novel empirical parameter, the Free Energy Deviation Index, is introduced and calibrated based on thermodynamic theory to quantitatively capture structural changes associated with different initial compaction water contents. Sensitivity analysis highlights the critical influence of soil structure on SWCC behavior. The proposed model demonstrates strong predictive performance against published experimental data and supports both forward estimation and inverse calibration, enabling more flexible and accurate integration of SWCC modeling into geotechnical engineering practice. Finally, a hybrid data-driven and physics-informed neural network (DD-PINN) framework is developed to characterize the rainfall-induced evolution of hydro-mechanical properties, namely coefficient of permeability and shear strength, which are critical inputs for the design and stability analysis of geo-infrastructures in unsaturated soils. In Stage I, an XGBoost–MLP proposed in Chapter 5 estimates the parameters of the FX-SWCC equation directly from basic index properties, eliminating labor-intensive laboratory testing and large labeled datasets. In Stage II, the predicted FX-SWCC parameters are embedded in a PINN model that enforces Richards' equation as a physical constraint to simulate the spatiotemporal evolution of matric suction and water content under transient boundary conditions. Shear strength is subsequently evaluated using the widely used Vanapalli et al. (1996) equation, while the coefficient of permeability is obtained via the Mualem (1976) relationship. By coupling a data-driven module that infers FX–SWCC parameters from basic index properties with a physics-informed PINN that enforces Richards' equation to predict the spatiotemporal evolution of SWCC, the proposed DD-PINN framework delivers mesh-free, physically consistent, and computationally efficient predictions of rainfall-induced changes in suction, coefficient of permeability, and shear strength, providing a unified tool for advanced slope-stability assessment and geo-hazard analysis. Recognizing that no single ML technique can address the breadth of tasks in unsaturated soil mechanics, this thesis examines black box, interpretable, data-driven, and physics-informed ML approaches and develops four practical frameworks that integrate these methods with the governing principles of unsaturated soil mechanics. This integration improves the estimation, prediction, and interpretation of hydromechanical behavior of unsaturated soils. The proposed frameworks provide efficient, interpretable, and physically grounded tools for estimating matric suction, modeling the SWCC, and predicting the rainfall-induced evolution of the coefficient of permeability and shear strength, thereby supporting rational design, performance evaluation, and hazard mitigation in geo-infrastructure founded on unsaturated soils.enUnsaturated soilMechanical propertiesMachine learning techniquesData-driven algorithmsPhysics-informed neural networkThesis