Deep learning assisted processing of synchrotron-based micro-CT imaging of fast, dynamic multiphase flow in porous media

The Australian Synchrotron micro-CT beamline was used to capture fast multiphase flow scansat 1s with 15-16s time lapse intervals at 3.6µm resolution. Two phase drainage flow experiments wereconducted using a custom-built rig with a water-wet and a mixed-wet Bentheimer sandstone sample toobserve the percolation of fluids in the pore spaces. During each injection cycle, fast dynamic batches of 50scans were taken resulting in large volumes of data. Conventional methods of image processing would notonly be time consuming but could also lead to misclassification of phases, especially due to the presence ofstreak artifacts in the datasets. A workflow was established using the Sci-Py library for data preprocessing(translation and rotation) of fast batch scans followed by a U-ResNet model for the accurate segmentationof the images. 25000 2D paired gray-scale and WEKA segmented 256x256 pixel patches were used fortraining the model, the mean Intersection-Over-Union (mIoU) showed an accuracy of 0.93 during the testingphase. While phase accuracy on unseen dataset was recorded on an average as 0.93, and mIoU of 0.96,visual examination indicated that streak artifacts had been minimized. U-ResNet model was furthercompared with watershed-based segmentation in terms of accuracy, speed and 2D Minkowski functionals.