Decoupled deep learning for geohazards mapping in oceanic deep drilling

• We map geohazards in no-well areas using seismic only with ≥90 % precision/recall. • A Fusion architecture decouples drilling geohazards via geological priors. • Achieves 0.904±0.113 F1-score in cross-validation using only 16% of the dataset. • Uncertainty-aware risk maps guide drill-site selection and optimize well design. Subsurface risk characterization in heterogeneous sedimentary basins remains a critical engineering challenge, particularly as conventional machine learning approaches often fail to capture complex, irreversible geological processes. We introduce a deep learning framework designed to address these limitations by integrating seismic attributes with knowledge-informed constraints through a decoupled architecture. The methodology addresses weak causal relationships by employing contrastive learning and a hybrid fusion regressor that reconciles high-frequency localized hazards with low-frequency basin-scale trends. Validated across 12 seismic surveys covering 4,100 km² of the Taranaki Basin, New Zealand (10.95 million seismic traces), our Fusion model trained on 299,169 derived data segments demonstrates superior performance, significantly outperforming conventional architectures based on BiGRU and Transformer models. Nine-fold cross-validation (using ∼16% of derived data, a subset randomly chosen to reduce computation from weeks to days) yields mean macro F1-score, precision, and recall of 0.904±0.113, 0.919±0.101, and 0.907±0.110, respectively, at 100m depth tolerance on test wells; on the full dataset, our approach achieves >98% precision and recall, maintaining robust generalization across unexplored offshore-to-inland transitions. These results demonstrate the practical engineering significance of our approach, establishing a scalable, data-driven method for pre-drill geohazard assessment that directly informs offshore resource management and drilling safety strategies.