Cross-well identification of gas hydrates based on well logging and machine learning: A case study from ODP Leg 204

Accurate identification of gas hydrate-bearing intervals is essential for marine energy resource evaluation and geohazard assessment. While machine learning (ML) approaches have shown promising performance in hydrate classification, most previous studies rely on pooled datasets or random sample splitting, which may overestimate predictive capability due to depth-dependent autocorrelation and limited domain separation. In this study, we investigate the cross-well generalization performance of ML models using well-logging data from three sites (1245, 1247, and 1250) of Ocean Drilling Program (ODP) Leg 204 at Hydrate Ridge. A strict Leave-One-Well-Out (LOWO) framework was adopted to enforce complete well-level separation and simulate deployment in unseen boreholes. Two gradient boosting models, XGBoost and CatBoost, were evaluated using seven logging-derived features, including elastic velocities, resistivity, density, porosity, gamma ray, and formation depth. Results show that meaningful cross-well discrimination is achievable (area under the curve = 0.83–0.97), although performance varies among wells due to geological heterogeneity and domain shift. Feature importance and SHAP analyses reveal that hydrate prediction is governed by a combination of thermobaric depth constraints and petrophysical indicators, particularly resistivity, bulk density, and elastic velocities. An exploratory cross-well regression experiment using continuous hydrate saturation estimates further indicates that relative structural trends in saturation are partially transferable across wells, whereas absolute magnitude prediction remains limited. These findings highlight the necessity of strict well-level validation and emphasize the influence of geological variability and label uncertainty on ML-based hydrate identification. The proposed framework provides a more realistic assessment of model transferability in heterogeneous marine hydrate systems.