Abstract

Geotechnical site characterization forms the cornerstone of reliable infrastructure design, yet traditional methods often struggle with the inherent spatial variability and uncertainty of subsurface conditions. This comprehensive review explores advanced data-driven machine learning (ML) frameworks that enhance soil stratification and parameter inference, leveraging in-situ tests such as cone penetration tests (CPT) and standard penetration tests (SPT). By integrating supervised, unsupervised, and ensemble learning paradigms, these frameworks address limitations in empirical correlations, offering improved accuracy in delineating soil layers and estimating key parameters like shear strength, compressibility, and modulus.
The paper begins with an overview of geotechnical challenges, including data scarcity, heterogeneity in urban environments like reclaimed lands in Singapore, and the need for probabilistic uncertainty quantification. Fundamental ML concepts are discussed, including algorithm selection, data preprocessing, and validation metrics such as root mean square error (RMSE) and coefficient of determination (R²). A detailed literature synthesis highlights the evolution from shallow learning models (e.g., support vector machines, random forests) to deep architectures (e.g., convolutional neural networks, long short-term memory networks) and hybrid approaches incorporating Bayesian inference for robust predictions.
Key advancements are examined in two core areas: soil stratification, where clustering techniques like k-means and density-based spatial clustering reveal depositional patterns; and parameter inference, where regression models predict engineering properties from fused datasets. Case studies from global sites, including seismic-prone regions in Asia, demonstrate practical applications, with performance benchmarks showing up to 30% improvement in accuracy over conventional methods. Challenges such as model interpretability, overfitting, and ethical considerations in high-stakes infrastructure are critically evaluated, alongside best practices for implementation.
Future directions emphasize integration with emerging technologies like digital twins, generative adversarial networks for synthetic data augmentation, and explainable AI to foster interdisciplinary collaboration. This review contributes a synthesized framework for practitioners and researchers, promoting data-centric geotechnics for safer, sustainable development in variable terrains.