Modeling subsurface geotechnical integrity via interpolated resistivity–chargeability and SPT datasets with machine learning: A case study from Perak, Malaysia

Abstract

Accurate subsurface assessment is critical for ensuring the integrity and longevity of engineered structures, especially in weathered terrains characterized by highly variable soil and rock properties. This challenge is particularly evident in the complex granitic terrain of Ipoh, Perak, Malaysia, where differential weathering and fracturing introduce substantial uncertainty into foundation design. Despite increasing demands for resilient infrastructure, conventional site investigation methods often fall short in capturing the spatial and lithological heterogeneity required for informed decision-making. This study addresses these limitations by integrating electrical resistivity tomography (ERT), induced polarization (IP), standard penetration testing (SPT), and supervised machine learning (ML) algorithms to enhance subsurface characterization and predictive modeling of chargeability. The combined ERT–IP inversion results, supported by borehole logs and SPT data, delineated four distinct lithological units across the study area. Four ML algorithms—SLR, KNN, SVM, and CatBoost—were trained to model relationships between geophysical and geotechnical parameters, with all achieving strong performance (training R² > 0.90; error metrics <10 %). Among all the evaluated models, CatBoost exhibited the strongest overall performance, achieving high predictive accuracy and maintaining consistent generalization across validation and test sets, although with mild overfitting observed. Among all the evaluated models, CatBoost exhibited the strongest overall performance, achieving high predictive accuracy and maintaining consistent generalization across validation and test sets, although with mild overfitting observed. From a geotechnical perspective, the topsoil/residual and highly weathered/fractured units were deemed unsuitable for heavy structural loads due to poor consolidation and water retention. In contrast, the relatively weathered bedrock unit, defined by resistivity values of 800 to >1000 Ωm and chargeability >18 msec, was identified as a suitable foundation material after appropriate soil and rock modifications. This research introduces a novel, noninvasive geophysical–geotechnical framework enhanced by ML, offering a cost-effective and scalable approach for subsurface evaluation. The methodology significantly reduces uncertainty in engineering assessments and supports safer, data-driven construction decisions, with potential applicability to other complex weathered terrains, subject to site-specific calibration.