Уear of publication: 2025

Fast multiphase processes in methane hydrate-bearing samples present major challenges for quantitative micro-CT experiments. These difficulties arise from the complex and evolving pore structure, the low contrast between solid and pore-space phases, and the large data volumes generated during dynamic synchrotron imaging. Traditional segmentation approaches are often manual and inaccurate under these conditions, limiting their utility for both quantitative analysis and real-time experimental feedback. These challenges extend further to mesh generation for subsequent numerical simulations. Multiphase datasets contain components with highly variable morphology and volume fractions. Phases with complex geometry and low volumetric content require fine elements to be accurately resolved, while applying such refinement uniformly to all phases results in oversampling and excessively large meshes. Given the scale of modern dynamic micro-CT datasets, mesh models should provide a balance between accuracy and computational efficiency. In this work, we utilize a two-step approach for automated segmentation of synchrotron dynamic micro-CT data of methane hydrate formation in coal media. The method combines a deep-learning 3D U-Net with Gaussian mixture clustering for robust separation of multiphase components. Meshing the hydrate phase is particularly challenging, as individual clusters may occupy only 1-2 voxels. To overcome this, we apply a temporal continuity strategy: the adapted mesh obtained at time step T is used as the initial mesh for step T+1, offering a consistent framework for monitoring the evolution of sub-resolution structures. Quantitative comparisons demonstrate that our adaptive meshing strategy achieves a significant reduction in computational cost while maintaining geometric accuracy for the dominant phases, compared to traditional voxel-based approaches.