Improving subsurface characterization in tropical granitic terrains using machine learning and deep learning-enhanced resistivity–chargeability modeling

Accurately characterizing subsurface lithological heterogeneity in tropical granitic terrains remains challenging because conventional electrical resistivity tomography (ERT) and induced polarization (IP) inversions often smooth abrupt boundaries, obscure weathered interfaces, and underrepresent clay–moisture contrasts. This study develops an integrated geophysical framework for approximating inversion-derived chargeability patterns using depth-referenced ERT–IP collocation, conservative outlier filtering, and stratified sampling to construct a robust dataset representative of residual soils, weathered and fractured granite, and moisture-bearing weak zones. Machine learning and deep learning (ML/DL) methods were applied to capture nonlinear resistivity–IP relationships governing lithological transitions. CatBoost achieved the strongest standalone performance (R 2 = 0.942 train; 0.945 test), marginally surpassing random forest, while the stacked ensemble achieved the highest predictive performance (test R 2 = 0.947). The one-dimensional convolutional neural network with global average pooling performed comparably to the artificial neural network while effectively representing depth-dependent resistivity–IP relationships. All ML/DL models yielded RMSE ≈ 0.76–1.09 msec (< 10%), confirming high predictive fidelity. ML/DL-derived and CatBoost-regressed 2D chargeability models reproduced inversion-defined structures and preserved the spatial expression of clay-rich horizons, fracture-controlled weak zones, and lithological interfaces. The results are relevant to geoengineering assessments, where relatively competent bedrock generally occurs below ~ 30–33 m and may provide favorable conditions for deep foundation development, subject to geotechnical verification. They also provide useful insights for hydrogeological evaluations by identifying weathered/fractured zones extending to ~ 41 m that may represent favorable groundwater targets. K-means clustering analysis (KMCA) further supported lithological discrimination by grouping materials into coherent residual-soil-to-weathered-granite domains. The integrated inversion–ML/DL–KMCA workflow provides a cost-effective and geologically consistent framework for approximating inversion-derived chargeability patterns and supporting subsurface characterization in complex crystalline terrains.