Quantitative analysis of dynamic CT imaging of methane-hydrate formation with a hybrid machine learning approach
Fast multi-phase processes in methane hydrate-bearing samples are challenging for micro-CT quantitative study because of complex tomographic data analysis involving time-consuming segmentation procedures. This is due to the sample multi-scale structure changing in time, low X-ray attenuation and phase contrast between solid and fluid materials, as well as large amount of data acquired during dynamic processes. We propose a hybrid approach for automatic segmentation of tomographic data from time-resolved imaging of methane gas-hydrate formation in sandy granular media. First, we use an optimized 3D U-net neural network to perform segmentation of mineral grains that are characterized by low contrast to the surrounding pore brine-saturated phases. Then, we perform statistical clustering based on the Gaussian mixture model for separating the pore-space phases that are characterized by gray-level instabilities caused by dynamic processes during hydrate formation. The proposed approach was used for segmenting several hundred gigabytes of data acquired during an in-situ tomographic experiment at a synchrotron. Automatic segmentation allowed for studying properties of the hydrate growth in pores, as well as dynamic processes such as the incremental pore-brine flow and redistribution.