Deep Learning Digital Rock Physics for Rock Petrophysical Properties

Abstract

This study introduces advanced optimization tools including machine learning models as cost-effective and efficient alternatives to traditional methods in the literature for estimating petrophysical properties in carbonate rock reservoirs. The models are trained and validated using experimental data and micro-CT images from carbonate rock samples.

First, we employ a stacked ensemble machine learning technique to predict carbonate rocks' porosity and absolute permeability using 2D slices of 3D micro-CT images. We leverage the water-shed technique to extract image pore regional properties, allowing for highly accurate porosity and absolute permeability predictions. By integrating predictions from multiple machine learning models into a single meta-learner model, our stacking ensemble learning approach accelerates prediction and improves the model's generalizability. This approach was tested on carbonate rock samples with different levels of heterogeneity, pore distributions, and absolute permeability, resulting in precise porosity and absolute permeability estimates.

Second, we demonstrate the power of convolutional neural networks and transfer learning in estimating rock porosity and absolute permeability from 2D slices of 3D micro-CT images. By leveraging pre-trained models from established literature, we replace the convolution layers with different depths and numbers of pre-trained convolution layers for feature extraction and achieve greater accuracy in predicting these critical rock properties. In this implementation, despite the generally expected reduction in computational time through the integration of a pretrained model, we observed an increase in computational requirements compared to the original convolutional neural network model. This deviation can be attributed to the complexity of our training dataset. Nevertheless, our results demonstrate the benefits of increased accuracy outweigh the additional computational time. This study reinforces the transformative potential of transfer learning in overcoming data limitations and enhancing prediction accuracy, making it an indispensable tool for geology and beyond.

Next, we extend the application of deep learning pipelines to determine multiphase flow petrophysical properties including relative permeability and capillary pressure. Laboratory assessments were performed on two-phase fluid flow within carbonate rock samples. Through unsteady-state core flooding experiments, we estimated both irreducible water saturation and residual oil saturation, and estimated the relative permeability endpoints. These parameters served as crucial boundary conditions for our analysis. We then utilized a well-established three-parameter model with mathematical modeling and optimization approaches to determine the variation of relative permeability with water saturation. The parameters of the three-parameter model were accurately determined, solving the fluid displacement equations and matching the average saturation measured experimentally at each time step within the Pyomo framework and interior point optimization (IPOPT) with the least square objective function. Moreover, we inferred the capillary pressure curves for the selected samples. We conducted a Sobol sensitivity analysis on the estimated parameters. This approach provides an alternative to traditional numerical simulation-based techniques for predicting critical reservoir properties.

Finally, we present a physics-informed machine learning approach for estimating relative permeability and capillary pressure curves from unsteady state coreflooding experimental data. This approach enables us to accurately solve governing equations and estimate parameters that describe the relative permeability function. By leveraging reinforcement learning, we can search for the best combinations of parameters to effectively define the variation of the relative permeability function within the physics-informed ML framework. Our method then utilizes these estimated parameters to solve for water saturation and predict capillary pressure curves for selected core samples. We also conducted a Sobol sensitivity analysis on the estimated parameters to ensure accuracy. Our proposed method offers a faster, more efficient, and more accurate alternative to traditional and time-consuming experimental and numerical simulation-based techniques. We have evaluated the accuracy of our approach in predicting relative permeability and capillary pressure curves in oil-water flow and compared it to Pyomo-based approach. In conclusion, our physics-informed machine learning approach presents a promising solution for accurately estimating relative permeability and capillary pressure curves from unsteady state coreflooding experimental data.

Our research has contributed to developing advanced tools that predict the petrophysical properties of reservoirs. These tools can potentially improve oil recovery and management in carbonate rock reservoirs and offer promising alternatives to traditional techniques.

Date of Award	11 Dec 2023
Original language	American English
Supervisor	HAMID Abderrahmane (Supervisor)