Machine learning-driven velocity–resistivity modeling: A novel explicit framework for near-surface characterization

Abstract

The integration of machine learning (ML) in geophysical investigations has become pivotal for resolving near-surface complexities, particularly in terrains with complex lithological heterogeneity. This study aims to develop and validate a novel ML-driven framework for jointly modeling seismic P-wave velocity (Vp) and resistivity relationships to improve subsurface characterization in tropical granitic terrains. Using collocated data from seismic refraction tomography (SRT) and electrical resistivity tomography (ERT) surveys across the South Penang Pluton, Malaysia, the method addresses traditional integration challenges by applying contour-based mesh interpolation to generate a densely aligned dataset—reducing manual bias, increasing data volume fivefold, and enhancing nonlinear predictive performance. Stratified binning ensured balanced lithological representation across training, validation, and test sets. Six ML models—simple linear regression (SLR), support vector machine (SVM), decision tree (DT), random forest (RF), gradient boost (GB), and artificial neural network (ANN)—were trained to predict Vp from resistivity. All models showed high predictive reliability (R² = 0.819–0.984, RMSE = 0.029–0.09, F1 = 0.878–0.910), outperforming previous regression-based approaches in the study area. ANN achieved the best performance, followed by SVM, both maintaining stability across all partitions. The framework integrates k-means clustering with internal validation and ML-predicted Vp–resistivity profiles to automate lithological classification into four distinct subsurface units. This integrated approach improves geophysical boundary resolution, preserves structural fidelity in zones of rapid geological transition, and enables scalable deployment in tropical weathered terrains where SRT coverage is often constrained. Ultimately, it advances geophysical site modeling from deterministic to data-driven prediction, offering cost-effective, transferable solutions for geotechnical, hydrogeological, and hazard-related applications in geologically complex settings.