Predicting soil properties for the purpose of site characterization using advanced Machine Learning approaches

Accurately predicting the spatial variation of undrained shear strength ( S u ) is critical for reliable geotechnical design, especially in heterogeneous subsurface conditions. This study presents a data-driven framework for probabilistic site characterization, predicting the spatial variability of S u using advanced machine learning techniques. Aimed at enhancing the reliability, reproducibility, and efficiency of geotechnical investigations, the methodology integrates both ensemble learning and deep learning models with domain-informed clustering strategies, following a systematic evaluation of multiple modeling approaches. The pipeline involves comprehensive data preprocessing, feature engineering, and the application of various supervised and unsupervised clustering techniques, to identify latent soil behavior groups. For single borehole prediction, underlying features of clusters using a k-means based k-Nearest Neighbours (kNN) approach are derived from clustered data, where each borehole is independently assigned to one of five clusters. Cluster-specific models are then separately trained, and ensemble regressors are used to predict the spatial variation of S u , utilizing observed data at each borehole to infer unobserved values. Furthermore, for the multi-borehole predictions, deep learning architectures, including Multimodal Transformer and Recurrent Neural Network (RNN) models, are considered and benchmarked against the ensemble learning approach. These models are evaluated in terms of predictive accuracy measured by root mean squared error (RMSE) and their ability to capture nonlinear relationships and spatial trends in the study area. The results demonstrate the effectiveness of combining local clustering with targeted model training in single borehole and consider deep learning approach for multi-borehole scenarios, providing a robust approach to modeling geotechnical variability for risk-aware site characterization. Comparative evaluation using a benchmark geotechnical dataset demonstrates that while ensemble learning provides reliable localized predictions under data-limited conditions, the transformer-based model significantly outperforms in capturing broader spatial trends and inter-borehole interactions. Furthermore, depth-wise LIME-based interpretability revealed that the transformer-based model adapts its predictions based on dominant geotechnical features at each depth range considering water content and geometry at shallow depths, and plasticity and elevation at greater depths.