A Generalized Borehole Autoregressive Neural Network for 3D Geological Modeling

Three-dimensional geological modeling is a fundamental technology for reconstructing subsurface geological structures and plays an important role in resource exploration, disaster prediction, and engineering construction. With increasing energy demand and growing environmental safety challenges, accurate characterization of the morphology and physical properties of subsurface strata has become essential for the efficient development of underground space. Machine learning-based three-dimensional geological modeling methods using borehole data reformulate the modeling process as a stratum classification task, thereby reducing manual intervention and improving the level of automation in geological modeling. In this process, the classification of stratigraphic spatial points is a key step, as its accuracy directly influences the quality of the resulting geological body model. However, traditional algorithms typically rely solely on spatial coordinate features to determine stratum affiliation. Such a single-feature-driven approach has limited capability in representing the true morphology of subsurface strata. To address this limitation, this paper proposes a stratum classification method based on Vertical Alignment–Horizontal Distance Weighting (VA-HDW), which is designed to capture spatial correlations between strata and boreholes. On this basis, a specialized neural network model, termed the Generalized Borehole Autoregressive Neural Network (GBARNN), is designed and trained to improve the classification performance of stratigraphic spatial points, thereby contributing to improved three-dimensional geological body modeling quality.