Machine Learning Approach for Prediction of Relative Permeability
Open Access
- Author:
- Yoga, Hanif
- Graduate Program:
- Energy and Mineral Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 02, 2022
- Committee Members:
- Mort Webster, Professor in Charge/Director of Graduate Studies
Russell Taylor Johns, Thesis Advisor/Co-Advisor
Hamid Emami-Meybodi, Committee Member
Yashar Mehmani, Committee Member - Keywords:
- machine learning
artificial intelligence
artificial neural network
relative permeability
hysteresis
equation of state
multiphase flow in porous media - Abstract:
- Hysteresis of transport properties like relative permeability (kr) can lead to computational problems and inaccuracies for various applications including CO2 sequestration and chemical enhanced oil recovery (EOR). Computational problems during multiphase numerical simulation include phase labeling issues and path dependencies that can create discontinuities. To mitigate hysteresis, modeling kr as a state function of key variables like wettability is a promising solution, because it honors changes in physical parameters and ensures a unique value of relative permeability for any set of input state parameters. In this research, we apply the state function concept to develop a physics-informed data-driven approach for predicting kr in the space of its state parameters. We extend the development of the relative permeability equation-of-state (kr-EoS) to create a predictive physics-based model using Artificial Neural Networks (ANN). We predict kr as a function of phase saturation and phase connectivity, as well as the specific saturation-connectivity path taken during the displacement, while maintaining other state parameters constant such as wettability, pore structure, and capillary number. We use numerical data generated from pore-network simulations (PNM) to test the predictive capability of the EoS. Physical limits within saturation-connectivity space are used to constrain the model and improve its predictability outside of the region of measured data. We find that the predicted relative permeabilities result in a smooth and physically consistent estimate. The results for a contact angle of 0° and 50° show that ANN can more accurately predict kr compared to using a higher-order polynomial response surface. With only a limited amount of drainage and imbibition data with an initial phase saturation greater than 0.7, a good prediction of kr is obtained from ANN for all other initial conditions, over the entire saturation-connectivity space. Finally for the nonwetting phase, we show good predictions of the specific path taken in the saturation-connectivity space along with the corresponding kr for any initial condition and flow direction, which makes the approach practical when phase connectivity information is limited. The path predictions for the wetting phase are not yet satisfactory, which shows a highly nonlinear path. This is left for future research. The research, however, does demonstrate the first application of a physics-informed data-driven approach for prediction of the nonwetting phase relative permeability using ANN.