Location of CO2 Leakage Plumes by Deep Learning Inversion of Gravity Data
Abstract
The US DOE National Risk Assessment Partnership (NRAP) is developing methods to evaluate the effectiveness of monitoring techniques to detect brine and CO2 leakage from legacy wells into underground sources of drinking water overlying a CO2 storage reservoir. As part of the NRAP Strategic Monitoring task group, we created 1,000 simulated aquifer impact data sets at 65 time-steps based on a hypothetic CO2 injection at the Kimberlina site in Southern San Joaquin Valley, which are used to generate 65,000 synthetic surface gravity data sets by forward models. One gravity data set consists of 451 data points on a rectangular grid that covers all likely plumes. Our previous study shows that the amplitude of gravity data may inform us the existence of CO2 leakage because it is sensitive to subsurface changes in CO2 saturation. But more information about leakage locations and CO2 saturation is needed for uncertainty reduction in risk assessment of CO2 leakage. We invert the gravity data for location and saturation of CO2 plumes with convolutional neural networks (CNN). The input of CNN is gravity data on the ground surface and the output is 3-D distribution of CO2 saturation. First, for input dimensionality reduction out of 451 data points per sample, we extract key features from the synthetic gravity data by downsampling and statistical analysis. Second, for output dimension reduction, we use principle component analysis through the method of snapshots and/or clustering analysis to extract the features (less than 65) of CO2 plumes with the plume location and CO2 saturation over 164,832 data points. 65,000 synthetic gravity data sets are used as training data. We adapt the objective function with regularization from geophysical inverse problems as the loss function of CNN. TensorFlow is used to generate the CNN model fitting the features from the gravity data onto the features of CO2 plumes with the plume location and CO2 saturation. It is expected that deep learning inversion of gravity data will have better resolution in locating large shallow CO2 plumes and quantify their CO2 saturation. Improvement may be achieved by optimizing feature extraction and loss function.
This work was performed under the auspices of the U.S. DOE by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-783480.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.S31E0577Y
- Keywords:
-
- 0599 General or miscellaneous;
- COMPUTATIONAL GEOPHYSICS;
- 0999 General or miscellaneous;
- EXPLORATION GEOPHYSICS;
- 7290 Computational seismology;
- SEISMOLOGY;
- 7299 General or miscellaneous;
- SEISMOLOGY