Using Unsupervised Machine Learning to Characterize Capillary Flow and Residual Trapping

Recent research has highlighted the impact mesoscale heterogeneity can have on larger‐scale multiphase fluid flow properties. However, currently, there is no consensus on how to quickly and reliably analyze coreflooding experimental data to gain insights into mesoscale capillary‐dominated flow behaviors and how rock petrophysical properties affect such behaviors. In this study, we combine a machine learning‐based clustering method with physics‐based hypothesis testing to analyze multistage steady‐state coreflooding experimental data. By examining five different sets of coreflooding data, we have found that under capillary‐dominated flow regimes, voxel‐level CO2 saturation fields are much more heterogeneous as capillary equilibrium is approached. CO2 saturation time series behaviors are distinctively different between high and low permeability voxels. However, the effect of mesoscale permeability greatly diminishes under viscous‐dominated flow regimes. By computing the maximum and optimal number of clusters on the CO2 saturation time series data, we are able to differentiate between these two flow regimes. Through this analysis method, it is now also possible to identify regions within a sandstone core with capillary heterogeneity trapping behaviors. It is confirmed that capillary heterogeneity trapping indeed occurs upstream of capillary barriers. Furthermore, we have discovered a new type of capillary heterogeneity trapping—one that occurs within the capillary barrier itself, which can cause parts of the core to achieve residual CO2 saturations much higher than expected based on existing models of residual gas trapping. Key Points An automated unsupervised machine learning method is developed to differentiate flow regimes and identify capillary heterogeneity trapping Capillary‐ and viscous‐dominated flow regimes can be differentiated by combining time series clustering and hypothesis testing This work shows the promise of machine learning methods for analyzing large data sets from coreflooding experiments