Deep Learning-Based Recognition and Quantitative Characterization of CO2 Hydrate Growth in Microfluidic Porous Media

The growth and morphological evolution of CO2 hydrates in marine sediments are recognized as highly significant for hydrate-based CO2 sequestration (HBCS). However, it remains challenging for their pore-scale processes to be directly characterized via macroscopic parameters. To ensure that issues such as weak boundaries, matrix texture interference, and internal pores in microfluidic CO2 hydrate images are addressed, an automated recognition and quantitative analysis method based on deep learning semantic segmentation was developed in this study. An image data set was compiled from microfluidic experiments, and four models (U-Net, Res-U-Net, Attention U-Net, and DeepLabv3+) were subsequently trained and compared. The results demonstrated that the hydrate regions were effectively identified by all four models. Among these, optimal performance was exhibited by the Attention U-Net model, with hydrate intersection over union (IoU) and Dice coefficients being recorded at 0.951 and 0.975, respectively. Furthermore, the optimal model was applied to continuous image sequences; consequently, the dynamic tracking of the CO2 hydrate growth process was achieved, and parameters, including hydrate saturation, internal pore ratio, and growth rate, were extracted. It was revealed that internal pores in the saline system are predominantly distributed within the hydrate envelope and that the inhibitory effect of the saline solution on hydrate growth is strengthened as the ion concentration is increased. At NaCl concentrations of 1, 3, and 5 wt %, the average growth rates were reduced by approximately 10.6%, 33.1%, and 70.4%, respectively. Among the various 3 wt % saline solutions tested, the lowest growth rate was observed in the NaCl system, whereas relatively high areal growth rates were maintained by the KCl, MgCl2, and CaCl2 systems. Thus, a reproducible pixel-level recognition and quantitative characterization approach for the microfluidic CO2 hydrate growth process is provided by our proposed method.