<p>Three-dimensional (3D) geological modeling is essential for urban planning and underground engineering, particularly in complex geological settings where the spatial distribution and stratigraphic sequence of geological bodies directly affect engineering safety. However, existing methods remain limitations in representing directional contacts, stratigraphic sequence, and thickness relationships of geological bodies, which may lead to discontinuity or inconsistence in geology models. To address these limitations, this study proposes a method based on digraph-based constraints, in which geological relationships are encoded using directed graphs and adjacency lists. The proposed constraints are integrated into a modeling framework that combines deep learning (DL) and multiple-point statistics (MPS), in which local pattern screening, global correction, and thickness constraints are used to guide sequential simulation and iterative optimization. The results show that the proposed method improves both geological consistency and quantitative reconstruction accuracy. In the simple sedimentary case, the average accuracy of borehole lithologies increased from 96.18% to 97.90%, while in the 3D Overthrust case it increased from 74.51% to 84.57%. Compared with the approach using stratigraphic sequence constraints, the proposed method reduced computational time by approximately 76.2%. In the Longzelu Station metro case, borehole-level validation at boreholes B-02, B-03, and B-05 yielded lithological matching accuracies of 83.0%, 80.9%, and 86.0%, respectively, demonstrating its practical applicability in engineering-scale modeling. Overall, the proposed method enhances the representation of directional geological relationships, improves model consistency, and provides a reliable framework for 3D geological modeling in complex underground engineering projects.</p>