A Three-Tier Evidence Hierarchy with Conditioned Generative Models for Discrete Fracture Network Reconstruction from Borehole Scanline Observations

Fracture networks govern the mechanical behaviour of rock masses, and their accurate characterisation underpins slope stability, excavation design, and rock mass classification. Direct field observations sample only a small part of the fracture system, so much of the geometry remains unknown. This study presents an evidence-guided framework for reconstructing two-dimensional discrete fracture networks from sparse borehole scanline observations, organised as a three-tier hierarchy based on observational certainty. Measured traces are retained as Tier 1 anchors under a tolerance-based containment criterion, indicated traces are extended by a family-conditioned length predictor built around an Enhanced Variational Autoencoder, and inferred traces are generated for unobserved regions by a sequential transformer that learns the typical spatial layout of each joint family. All candidates pass through a multi-objective scorer that enforces seven geological constraints. The framework was validated on nine fully mapped bench faces of an open-pit mine, comprising 512 traces across three systematic joint families and random joints, at borehole spacings of 3 to 11 m. Relative to a Monte Carlo baseline, it reduced held-out per-family orientation error for all three joint families and improved the overall agreement with the mapped orientation and length distributions by about a third. Leave-one-bench-face-out testing confirmed that this performance reflects the framework rather than the particular faces held out. Reconstruction quality depended on a trade-off in observation spacing, with intermediate layouts giving the best overall balance across orientation accuracy, spatial fidelity, and network topology for the studied rock mass.