Deep learning for automated characterization of pore-scale wettability
Introduction
Coupled physical mechanisms resulting from viscous, capillary, and surface forces make the description of multiphase flow in porous media challenging. At the pore scale, the interplay of pore network geometry, connectivity, and surface wettability play an important role in the distribution of fluid phases. Improved understanding of subsurface fluid flows, and specifically the pore scale, is required to facilitate critical processes in the context of water and energy resources, including recovery of spilled hydrocarbons, enhanced oil recovery, and geological carbon sequestration.
Critical understanding of multiphase flow has been enabled by microfluidic devices known as micromodels, advanced imaging technologies, and improved computational power (Armstrong et al., 2016). Micromodels contain microlevel representations of rock structure and are often made of silicon, e.g., Buchgraber et al. (2012a), glass. or HDMS. The strength of microfluidics for characterization of multiphase flow effects in porous media include reproducibility and control over pore structures, optical access to pore-scale events, high resolution images, and pore-scale controls of fluid flow. In the context of subsurface water and energy resources, micromodels reveal mechanisms controlling multiphase flow and transport phenomena in the pore space of geological porous media (Sinton, 2014). For instance, understanding fluid flow through dual-porosity carbonate rock is complex due to its heterogeneous pore network across multiple length scales. A typical dual-porosity carbonate pore network consists of a wide size range of pores (1–200 µm) and minerals mixed with micro- and macroscale fractures and particles of varying size (Yun et al., 2017).
Micromodels are often criticized for their lack of reproduction of solid/fluid chemistry whereas the effect of wettability (i.e., the affinity of fluid for a pore wall) on fluid movement can be studied and visualized particularly well in the pore space of micromodels (Zhao et al., 2016). A typical etched-silicon micromodel has uniform silicon dioxide pore surfaces after manufacture. Accordingly, recent papers have emphasized the functionalization of micromodels with clay minerals and calcite (Song, Ogunbanwo, Steinsbø, Fernø, Kovscek, 2018, Song, Kovscek, 2015, Yun, Chang, Cogswell, Eichmann, Gizzatov, Thomas, Al-Hazza, Abdel-Fattah, Wang, 2020) as well as construction of micromodels out of calcite (Song et al., 2014). Other studies have focused on the understanding of how chemicals alter the arrangement of fluids at pore scale (Howe, Clarke, Mitchell, Staniland, Hawkes, Whalan, 2015, Yun, Chang, Cogswell, Eichmann, Gizzatov, Thomas, Al-Hazza, Abdel-Fattah, Wang, 2020).
The accurate and rapid characterization of trapping processes, the arrangement of fluids at pore scale, and the impact of trapping on multiphase displacement in micromodels remains a challenge. In short, it is not yet currently feasible, to perform quick and thorough interrogation of a large number of sample images of rock and its pore space at high resolution over a large Field of View (FOV). Thus, full characterization of the interactions between aqueous phases, organic phases, and rock is limited. We believe that pore-to-Darcy scale observation of displacement processes in complex and realistic porous media (e.g., sandstone-type and carbonate-type systems) can be achieved by fast monitoring using a large number of microscopic images of pores existing in porous media.
Motivated by advances in related fields that use microfluidics (Riordon et al., 2019), we explore machine learning for the rapid and accurate categorization of pore-scale images. We place special emphasis on the fluids that wet pore walls because of the importance of wettability to multiphase flow (Blunt et al., 2013) and the various industrially-relevant methods employed to change wettability at the pore scale (Lu, Goudarzi, Chen, Kim, Delshad, Mohanty, Sepehrnoori, Weerasooriya, Pope, 2014, Hirasaki, Miller, Puerto, 2008) Image analysis to quantify the volume fraction of pore space filled by aqueous, organic, and solid phases is relatively easy to automate, but the interpretation of the fluids that wet solid surfaces remains painstaking and laborious. Hence, the main objective of this manuscript is to explore the application of deep learning to enable automatic categorization of wettability using pore-scale images. We contribute here to understanding of multiphase flow through aiding the experimentalist. Toward this purpose, this paper proceeds with a literature review including wettability, microfluidics for multiphase flow applications, and applications of machine learning to porous media images. Then the methodologies for image acquisition and the training of deep learning models to classify wettability using Fully Connected Layer Networks (FCN) and Convolutional Neural Networks (ConvNet) are discussed. Image classification results follow including a saliency analysis that illustrates qualitatively and quantitatively the role of each pixel on image classification.
Section snippets
Literature review
This section discusses the importance of understanding the impact of wettability of the porous system where multiple phases (oil, water, and gas) coexist. Micromodels as a means of understanding wettability and flow mechanisms are also discussed. Then, recent studies employing machine learning for accelerating data analysis are presented.
Data acquisition
An experimental procedure to fabricate microfluidic devices with strong oil- and water-wet surface wetting conditions is presented. Importantly, Laser Scanning Fluoresence Microscopy (LSFM) is used to collect many pore-scale images. Image processing, post image acquisition, and details of the data structure for images are also presented in this section to document the use of microfluidic image data and deep-learning.
Training
Spatial resolution is important for wettability classification. Hence, we explain the selection criteria for the 6400 images with different FOVs for all classes. This section discusses our methodology using fully connected neural networks and convolutional neural network models in detail. Its also details the evaluation criteria.
Results and discussion
The wettability and morphology of pores strongly influences the ganglia shape, and pore-scale configuration of oil and water phases. This section presents the predictions of wettability (strong oil-wetting or strong water-wetting) and illustrates the interplay between the oil phase, water phase, and grain surfaces.
Summary and future work
This work presented and evaluated a methodology for automated classification of images from microfluidic devices (micromodels) with water- and oil-wet surface wettability. Fluid injection experiments using micromodels with two different pore network patterns were conducted to prepare pore-level images that were used as a data-set to train two deep-learning algorithms. The first is a three-layer fully connected neural network (FCN) and the second is a convolutional neural network (ConvNet). Both
CRediT authorship contribution statement
Wonjin Yun: Conceptualization, Methodology, Software, Writing - original draft. Yimin Liu: Conceptualization, Methodology, Software, Writing - original draft. Anthony R. Kovscek: Conceptualization, Supervision, Funding acquisition, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
The support of the SUPRI-A Industrial Affiliates is acknowledged gratefully.
References (40)
- et al.
The effect of wettability on capillary trapping in carbonates
Adv. Water Resour.
(2016) - et al.
Pore-scale contact angle measurements at reservoir conditions using x-ray microtomography
Adv. Water Resour.
(2014) - et al.
Pore-scale imaging and modelling
Adv. Water Resour.
(2013) - et al.
Creation of a dual-porosity micromodel for pore-level visualization of multiphase flow
J. Petrol. Sci. Eng.
(2012) - et al.
Visualising surfactant enhanced oil recovery
Colloids Surf., A
(2015) - et al.
Enhanced oil recovery from high-temperature, high-salinity naturally fractured carbonate reservoirs by surfactant flood
J. Petrol. Sci. Eng.
(2014) - et al.
Automated advanced image processing for micromodel flow experiments; an application using labview
J. Petrol. Sci. Eng.
(2018) - et al.
Direct pore-scale visualization of interactions between different crude oils and low salinity brine
J. Petrol. Sci. Eng.
(2018) - et al.
Deep learning with microfluidics for biotechnology
Trends Biotechnol.
(2019) - et al.
Direct visualization of pore-scale fines migration and formation damage during low-salinity waterflooding
J. Nat. Gas Sci. Eng.
(2016)
In situ characterization of mixed-wettability in a reservoir rock at subsurface conditions
Sci. Rep.
Deep Learning Convolutional Neural Networks to Predict Porous Media Properties
SPE Asia Pacific Oil and Gas Conference and Exhibition
Beyond Darcy’s law: the role of phase topology and ganglion dynamics for two-fluid flow
Phys. Rev. E
Effect of surface wettability on immiscible displacement in a microfluidic porous media
Energies
A study of microscale gas trapping using etched silicon micromodels
Transp. Porous Media
Deep Learning
Silane modification of glass and silica surfaces to obtain equally oil-wet surfaces in glass-covered silicon micromodel applications
Water Resour. Res.
Deep residual learning for image recognition
Proceedings of the IEEE conference on computer vision and pattern recognition
Recent advances in surfactant eor
SPE Annual Technical Conference and Exhibition
Effects of pore-Scale disorder on fluid displacement in partially-Wettable porous media
Sci. Rep.
Cited by (11)
Machine learning - based shale wettability prediction: Implications for H<inf>2</inf>, CH<inf>4</inf> and CO<inf>2</inf> geo-storage
2024, International Journal of Hydrogen EnergyMicrofluidic droplet detection via region-based and single-pass convolutional neural networks with comparison to conventional image analysis methodologies
2022, Machine Learning with ApplicationsSpecial issue on machine learning for water resources and subsurface systems
2021, Advances in Water ResourcesMicrofluidic devices, materials, and recent progress for petroleum applications: A review
2024, Canadian Journal of Chemical EngineeringMultivariate Geostatistical Simulation and Deep Q-Learning to Optimize Mining Decisions
2023, Mathematical Geosciences