The top figure shows a geologic area and the seismic survey over that area. Red symbols represent sources, and yellow symbols represent receivers needed to conduct that survey. There are two wells: the left well is injecting CO2 in this area and the right well is monitoring the underground CO2 flow. The bottom figures show the CO2 plume overlaid over two different permeability models. Since permeability is the parameter which defines the ease of flow of any fluid in the subsurface, we can see shape of CO2 plumes are different in the bottom left and right figures. This figure also signifies the importance of monitoring CO2 storage projects because in the bottom right figure we can see that CO2 plume is almost breaching the storage complex which is undesirable for the regulators.
As greenhouse gases accumulate in the Earth’s atmosphere, scientists are developing technologies to pull billions of tons of carbon dioxide (CO2) from the air and inject it deep underground.
The idea isn’t new. In the 1970s, Italian physicist Cesare Marchetti suggested that the carbon dioxide polluting the air and warming the planet could be stored underground. The reality of how to do it cost-effectively and safely has challenged scientists for decades.
Geologic carbon storage — the subterranean storage of CO2 — comes with significant challenges, most importantly, how to avoid fracturing underground rock layers and letting gas escape into the atmosphere. Carbon, a gas, can behave erratically or leak whenever it’s stored in a compressed space, making areas geologically unstable and potentially causing legal headaches for corporations that invest in it. This uncertainty, coupled with the expense of the carbon capture process and its infrastructure, means the industry needs reliable predictions to justify it.
Georgia Tech researcher and Georgia Research Alliance Eminent Scholar Felix J. Herrmann has an answer. His lab, Seismic Laboratory for Imaging and Modeling (SLIM), uses advanced artificial intelligence (AI) techniques to create algorithms that monitor and optimize carbon storage. The algorithms work as “digital twins,” or digital replicas of underground systems, facilitating the safe, efficient storage of CO2 underground.
“The trick is you want the carbon to stay put — to avoid the risk of, say, triggering an earthquake or the carbon leaking out,” said Herrmann, professor in the School of Earth and Atmospheric Sciences and the School of Electrical and Computer Engineering. “We’re developing a digital twin that allows us to monitor and control what is happening underground.”
Predicting the Best Place
Waveform Variational Inference via Subsurface Extensions with Refinements (WISER) is an algorithm that uses sound waves to analyze underground structures. WISER runs on AI, enabling it to work more efficiently than most algorithms while remaining computationally feasible. To improve accuracy, WISER makes small adjustments using sound wave physics to show how fast sound travels through different materials and where there’s variation in underground layers. This helps to create detailed, reliable images of underground areas for better predictions of carbon storage.
WISER allows researchers to work with uncertainties, which is vital for understanding the risk of these underground storage projects.
Scaling the Algorithm
While Herrmann’s lab has been working to apply neural networks to seismic imaging for years now, WISER required them to increase the networks’ scale. Making multiple predictions is a much larger problem that requires a bigger, more potent network, but these types of neural networks only run on graphics processing units (GPUs), which are known for speed but are limited in memory.
To optimize the GPU, Rafael Orozco, a computational science and engineering Ph.D. student, created a new type of neural network that can train with very little memory. This open-source package, InvertibleNetworks, enables the network to train on very large inputs and create multiple output images conditioned on the observed seismic data.
WISER’s fundamental innovation is for the lab’s next concept: creating digital twins for carbon storage. These twins can act as monitoring systems to optimize and mitigate risks of carbon storage projects.