Application of machine learning for predicting unfrozen water content in frozen soils: A review

Accurate estimation of unfrozen water content (UWC) in frozen soils is critical for understanding environmental and geotechnical processes in cold regions. Experimental determination, although accurate, is typically labor-intensive, costly, and limited in its application. Moreover, conventional modeling approaches are often constrained by simplified assumptions and reliance on specific soil parameters, thereby limiting their generalizability and predictive accuracy. Consequently, machine learning (ML) approaches have emerged as promising alternatives because of their computational efficiency and capability to handle complex nonlinear relationships. This review systematically evaluates the application of ML for UWC prediction in frozen soils, tracing the field's evolution from initial exploratory or standard models to sophisticated ensemble algorithms. A comprehensive bibliometric and thematic analysis of existing literature highlights significant methodological advancements, particularly the shift toward physics-informed ML (PIML) models. Despite these advancements, several critical challenges persist, including data scarcity and fragmentation, limited generalizability, and insufficient emphasis on interpretability and uncertainty quantification. To address these barriers and facilitate broader adoption of ML in practical engineering contexts, the review emphasizes the necessity of developing large-scale, standardized benchmark datasets, advancing PIML frameworks, and integrating robust interpretability and uncertainty analysis into standard ML workflows. By systematically addressing these issues, ML can evolve from a promising tool into an indispensable, trusted component integral to modern cold regions science and engineering. • Using ML algorithms to predict UWC is a developing research frontier. • Major challenges include data scarcity, poor generalization and interpretability. • A general end-to-end modeling framework for UWC estimation is proposed.