Soft-computing framework for 3D subsurface geospatial modeling using point-cloud-driven ensemble machine learning and deep learning

초록

In modern geotechnical practice, high resolution 3D subsurface models are essential for understanding complex stratigraphy and supporting design of resilient infrastructure. We present a data-driven framework that converts sparse 2D borehole logs into detailed 3D voxel models via an ensemble learning approach. A large synthetic geotechnical dataset, generated by parametrically sampling layer thicknesses and lithologies from real borehole statistics, was used to assess model performance under varying borehole density and geological variability. Within this framework we compare machine learning (ML) models (ANN, RF, SVM, etc.) against 3D deep learning (DL) architectures (PointNet, VoxelNet, SECOND, PointPillars) for subsurface layer classification and property regression. Consistent with prior geotechnical ML studies, it was observed that among the ML models an artificial neural network achieved the highest accuracy. Among the 3D deep learners, a VoxelNet-style volumetric network was most effective particularly in cases of high spatial heterogeneity. These results highlight the value of volumetric deep learning; recent studies have suggested that deep-network approaches can complement, and in some settings outperform, classical geostatistical interpolation in 3D ground modeling. Finally, our high-fidelity 3D models have direct engineering relevance. Accurate voxelized stratigraphy aids site-specific analyses and hazard mapping—e.g. 3D depth-to-bedrock models inform tunnel and foundation design, risk assessment, and smart-city planning. Overall, this work demonstrates how integrating modern 3D DL into geospatial ground modeling can improve predictive accuracy relative to traditional ML, paving the way for advanced geotechnical decision support. © 2026 Elsevier B.V.

키워드