Physics-Informed Deep Learning For Prediction of CO2 Storage Site Response
Open Access
- Author:
- Prathipati, Sumedha
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 18, 2021
- Committee Members:
- Chitaranjan Das, Program Head/Chair
Daniel Kifer, Thesis Advisor/Co-Advisor
C Lee Giles, Committee Member
Parisa Shokouhi, Committee Member - Keywords:
- CO2 storage
Site response forecasting
Physics-informed deep learning
Long Short-Term Memory (LSTM) - Abstract:
- The accurate prediction of CO2 saturation and pressure is crucial for effective management of carbon storage sites. As the extensive pre-injection tests of such carbon sites are not strategically feasible, the reservoir behavior can be replicated by numerically simulating the CO2 flow. Prior knowledge of the carbon storage site response to injection of CO2 and the underlying geological formations of the reservoir are required in order to develop an economic strategy to monitor the storage projects. In recent years, these accurate simulations have been modeled by data-driven machine learning methods. Although such models are computationally inexpensive, these may overfit the data and not take the underlying physical laws into account. Here, a physics-informed deep learning method which incorporates flow equations is proposed to predict a storage site response to injection of CO2 gas. This approach is demonstrated using a 3D synthetic dataset. The model approximates the evolution of CO2 saturation, pressure and water production rate with respect to space and time, given the initial permeability, porosity and CO2 injection rate. First, a data-driven Long Short-Term Memory (LSTM) model is developed as the baseline. The physics-informed LSTM model is built on this by adding constraints to the cost function which are defined by the governing physics equations for a two-phase flow system. The proposed approach can be incorporated in carbon storage management to accurately predict the storage site response to CO2 injection.