Gas hydrate (GH) resources in the Ulleung Basin hold promise for enhancing South Korea’s energy security; however, their commercial development remains constrained by technical uncertainties. This study presents a hybrid artificial intelligence (AI) framework combining supervised and unsupervised learning to improve the interpretation of GH behavior during laboratory depressurization experiments. A convolutional neural network (CNN) is trained to predict three‐phase saturations—water, GH, and gas—using X‐ray computed tomography (CT) images. Physically consistent labels were generated using a material balance equation incorporating phase‐specific densities to ensure saturation summation constraints. Latent features extracted from the CNN’s flattened layer were visualized using t‐distributed stochastic neighbor embedding (t‐SNE) to reveal distinct clusters corresponding to GH formation and dissociation stages. Compared to t‐SNE applied directly to raw CT images, the CNN‐based embeddings demonstrated markedly improved cluster compactness and separation. This improvement was quantified using the simplified Davies–Bouldin and within (S‐DBW)‐cluster scatter metrics, which demonstrated enhanced clustering performance—showing a 37.5% reduction in the average S‐DBW value and a 56.0% reduction in standard deviation compared to the base case. Sensitivity analysis further confirmed the robustness of the CNN‐based visualization across a wide range of t‐SNE perplexity settings. The resulting cluster distributions aligned well with known physical transitions in GH systems, such as the dissociation threshold near 16 MPa and corresponding shifts in phase saturations. These findings demonstrate the CNN’s ability to extract meaningful, physically relevant features from high‐dimensional image data, enabling more interpretable and reliable analysis of multiphase systems. This hybrid framework offers not only improved predictive accuracy but also a robust and interpretable tool for analyzing GH experimental data. The methodology is readily extendable to other geoscience applications involving complex pore‐scale imaging and fluid behavior, providing a novel pathway for integrating deep learning with domain expertise in subsurface energy research.
Kim et al. (Thu,) studied this question.