This study investigates the formation mechanism of natural gas hydrates in the deep-water Qiongdongnan Basin, with particular emphasis on the controlling role of sandy reservoirs in the hydrate accumulation system. On the basis of geological observations of sand bodies inside and outside the gas hydrate stability zone (GHSZ), including locally continuous hydrate-bearing sand bodies within the stability zone, the sealing effect of overlying mudstones, and sandy reservoirs near the base of the stability zone that provide pathways for the migration of underlying free gas, a new conceptual model for hydrate accumulation is proposed. A Transformer-encoder-based artificial intelligence algorithm was applied to three-dimensional seismic data and associated well-log data to predict the spatial distribution of sandy reservoirs and gas hydrate enrichment zones in the study area. The prediction results indicate high-saturation hydrate enrichment within high-quality sandy reservoirs inside the stability zone, and also suggest the sealing efficiency of overlying mudstones and the possible connectivity between sand bodies beneath the stability zone and deeper gas sources. This study supports the proposed accumulation mechanism and illustrates the application of artificial intelligence methods for hydrate reservoir prediction under complex geological conditions, providing an integrated interpretation and technical reference for selecting natural gas hydrate exploration targets in the deep-water northern South China Sea.

Submarine natural gas hydrate accumulation requires deep methane to enter the gas hydrate stability zone, become immobilized within a reservoir, and remain confined beneath a caprock. The long-term coupling among reservoir supply, caprock retention, and hydrate self-sealing remains poorly resolved. We developed a mass-conservative synthetic two-dimensional Darcy flow and reactive-transport model that couples a quasi-steady potential field, mobile methane transport, supply- and pore-capacity-limited hydrate formation, and saturation-dependent permeability reduction. Single-factor, full-factorial, global-sensitivity, numerical-convergence, and long-term experiments showed that reservoir and caprock permeability both controlled methane supply; hydrate mass was more sensitive to caprock permeability (PRCC = 0.919) than to reservoir permeability (PRCC = 0.738), and the reservoir response plateaued above approximately 100 mD. Caprock permeability and thickness jointly regulate flux and leakage, with maximum conversion occurring at intermediate caprock permeability rather than at maximum flux. In the 30 kyr reference case, self-sealing reduced source-zone flow by 7.14%. In the 500 kyr extension, hydrate-mass growth and source-flow change first remained simultaneously below 1% over the preceding 10 kyr at approximately 338 kyr. The parameter combinations maximizing hydrate mass, conversion efficiency, and sustained accumulation differed, and, at 30 kyr under the stated thresholds, only 8 of 81 combinations jointly achieved high conversion, elevated reservoir saturation, and low leakage. Favorable accumulation therefore reflects dynamic matching among supply, retention, and self-sealing under specified boundary conditions, rather than an isolated extreme in any single property.

Clay-surface conduction in fine-grained, weakly consolidated marine sediments, together with the scale mismatch between core and logging observations, often causes resistivity- and acoustic-based estimates of gas hydrate saturation to diverge. For Well GMGS5-W08, we develop a core-constrained, scale-matched electroacoustic fusion method (CC-SMEF) for gas hydrate saturation estimation. The method derives a low-frequency clay background from gamma-ray and core clay data. An optimizable vertical window then summarizes the logs around each core reference depth as robust neighborhood features. A low-dimensional, regularized kernel map integrates deep resistivity, P-wave velocity, gamma ray, and the clay background on this common representation scale. Chlorinity-derived saturation provides the geochemical reference, and the model was compared with the Archie and Simandoux methods at all reference points and with the isotropic three-phase acoustic method at available matched depths. Pointwise leave-one-out evaluation indicated lower observed error than Archie, although the uncertainty interval included no improvement and explanatory capability remained moderate. The fixed-parameter ablation is exploratory. Under the stricter four-block protocol, saturation-model tuning and fitting were restricted to training blocks while the clay background remained calibrated from all 22 core clay measurements; CC-SMEF RMSE was 17.89 percentage points versus 15.87 for Archie, so superiority in held-out-depth prediction was not established. CC-SMEF is therefore interpreted as a local interpolation framework for the represented W08 intervals; application elsewhere requires independent calibration and validation.