A physics-informed deep learning method for prediction of CO2 storage site response

Abstract:

We present a physics-informed deep learning method for estimating pressure, gas saturation and water extraction rate of CO2 storage sites. Developing a cost-effective strategy to operate CO2 storage sites requires knowledge and understanding of the site response to CO2 injection given the uncertainties in geological characterization. Compared to the petroleum industry, the experience related to operating and regulating commercial-scale CO2 storage sites is very limited. In addition, pre-injection field tests are not economically viable or strategically feasible. In order to compensate for the lack of sufficient field data and experience, high-fidelity numerical simulations and machine learning may be combined to create the possibility to quickly explore and test CO2 storage site behavior in various operational scenarios. This will provide the stakeholders with the necessary tools to observe the impact of different operational parameters on a variety of outcomes (e.g., pressure, storage capacity and recovery) over time. In this study, we use deep learning to approximate the site response (pressure, CO2 saturation and water extraction rate) over time, given initial values of porosity, permeability and injection rate. The model is trained, validated and tested using several synthetic datasets of increasing complexity in terms of geometry, geology, and heterogeneity. We first compare the prediction results obtained using Long Short-term Memory (LSTM), Convolutional Neural Networks (CNN), Multilayer Perceptron (MLP), among others. Next, we use simplified physics-based partial differential equations (conservation of mass coupled with Darcy’s law for a two-phase system) to modify the loss term in the MLP model. Implementing the physical knowledge is shown to improve the model performance. The proposed modeling approach can be integrated in CO2 storage management to predict the critical site response parameters for a range of relevant input parameters.