Detection of CO2 Leakage Plumes by Deep Learning Inversion of Gravity Data with U-Net

Abstract:

The US DOE National Risk Assessment Partnership (NRAP) is developing methods to evaluate the effectiveness of monitoring techniques to detect CO₂ leakage from legacy wells into shallow aquifers. The gravity method is sensitive to changes in CO₂ saturation incurred by CO₂ leakage. The conventional deterministic inversion of gravity data produces CO₂ plume images with poor resolution. We conducted deep learning inversion of gravity data for better resolution of CO₂ plumes using U-shaped convolutional neural networks (U-Net). A U-Net consists of a contraction path that determines if CO₂ plumes are present in the subsurface, plus a symmetric expansive path that quantifies the location and saturation of the CO₂ plumes. We used the U-Net to predict a subsurface cross-section CO₂ plume image from the gravity image on the ground surface.

We generated two sets of CO₂ leakage plume samples to train the U-Net. The first set consists only of single plumes with 720 samples at six possible depths, 12 saturation levels and 10 plume volumes. The synthetic gravity data corresponding to each plume sample is generated by a gravity forward model. The U-Net can predict an image of a single CO₂ plume accurately from gravity data. The second dataset consists of over 200,0000 double-plume samples. Due to limited computer memory, we randomly down-selected about 30,000 samples from the second dataset for training the U-Net. Detection of two plumes by the deep learning inversion of gravity data with a U-Net requires hyperparameter optimization, custom loss function and mitigation of vanishing gradients. The U-Net can locate two plumes in most of the test samples, but improvements are desired to better resolve the plume size and CO₂ saturation. A significant advantage of the deep learning inversion over the conventional inversion is that small and deep CO₂ plumes can be detected by a U-Net. Our future work will focus on training the U-Net with physics-based simulated CO₂ leakage data. Deep learning inversion has the potential to provide instant interpretation of monitoring signals and facilitates near real-time monitoring of geologic carbon sequestration.

This work was performed under the auspices of the U.S. DOE by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-812923.