Jyoti Behura, Colorado School Of Mines
A critical goal of any CCUS project is to map the CO₂ plume as a function of space and time. Here, I present a methodology to map the CO₂ saturation and pore-pressure from seismic data using deep learning. Conventional time-lapse processing of seismic data yields a qualitative image of the subsurface which needs further interpretation or analysis to extract quantitative information out of it. Furthermore, the process is time-consuming, expensive, and requires substantial manual effort.
Here, I present an alternative approach whereby the raw seismic data is directly converted to a petrophysical image of the subsurface. We generate depth images of changes in CO₂ saturation and pore-pressure. This approach is cost-effective, can be performed in real-time, yields quantitative information about the petrophysical changes.
Seismic data is input to the deep learning network and pore-pressure and saturation make the network outputs. I demonstrate the presented methodology on synthetic data acquired over the Kimberlina geologic model, which has been extensively studied. In order to be close to reality, however, I perform elastic simulations in an attenuation medium. Moreover, I also employ multi-component data so as to take advantage of all the information contained in the seismic data. To make the method computationally efficient, I transform the input seismic data to the frequency domain, whereby the dimensions of the input data are more amenable to convolutional neural networks. Another key assumption I make is that the medium is locally one-dimensional, which is true for almost all CCUS projects around the world. This assumption also makes the neural network efficient as well as provides additional training data.
The neural network is trained on data generated from seismic data acquired before CO₂ injection and 1-year after initiation of injection. None of the data from any other subsequent surveys is employed in the training. The trained model is thereafter applied to time-lapse data from subsequent surveys. The results show that both CO₂ saturation and pore-pressure perturbations are faithfully reproduced (qualitatively as well as quantitatively) by the deep learning model for seismic surveys as far as 20 years after injection. The method is computationally efficient, can operate on sparse data, as well as perform targeted imaging.