A team of scientists has created a new AI-based tool to help block greenhouse gases like CO2 in porous rock formations faster and more accurately than ever.
Carbon capture technology, also known as carbon sequestration, is a climate change mitigation method that redirects CO2 issued by underground power plants. In doing so, scientists must avoid excessive pressure build-up caused by CO injection2 in rock, which can fracture geological formations and release carbon into the aquifers above the site, or even into the atmosphere.
A new neural operator architecture called U-FNO simulates pressure levels during carbon storage in a split second, doubling the accuracy on certain tasks, helping scientists find optimal injection speeds and sites. It was unveiled this week in a study published in Advances in water resourceswith co-authors from Stanford University, the California Institute of Technology, Purdue University, and NVIDIA.
Carbon capture and storage is one of the few methods that industries such as refining, cement and steel could use to decarbonise and meet emission reduction targets. Over one hundred carbon capture and storage facilities are under construction around the world.
U-FNO will be used to accelerate carbon storage predictions for ExxonMobil, which funded the study.
“Tank simulators are intensive computer models that engineers and scientists use to study multiphase flows and other complex physical phenomena in the earth’s underground geology,” said James V. White, head of underground carbon storage at ExxonMobil. “Machine learning techniques such as those used in this work provide a solid path to quantify uncertainties in large-scale underground flow models such as carbon capture and sequestration, and ultimately facilitate better decision making.”
How carbon storage scientists use machine learning
Scientists use carbon storage simulations to select the correct injection sites and rates, control pressure build-up, maximize storage efficiency, and ensure that injection activity does not disrupt rock formation. For a successful storage project, it is also important to understand the carbon dioxide plume, the spread of CO2 across the land.
Traditional simulators for carbon sequestration are time-consuming and computationally expensive. Machine learning models provide similar levels of accuracy by dramatically reducing the time and cost required.
Based on the U-Net neural network and Fourier’s neural operator architecture, known as the FNO, U-FNO provides more accurate predictions of gas saturation and pressure build-up. Compared to using a state-of-the-art convolutional neural network for activity, U-FNO is twice as accurate while requiring only a third of the training data.
“Our machine learning method for scientific modeling is fundamentally different from standard neural networks, where we typically work with fixed resolution images,” said Anima Anandkumar, co-author of the paper, director of machine learning research at NVIDIA and Bren professor at Computing + Department of Mathematics at Caltech. “In scientific modeling, we have different resolutions depending on how and where we sample. Our model can generalize well on different resolutions without the need to retrain, obtaining huge speed increases ”.
The trained U-FNO models are available in a web application to provide real-time forecasts for carbon storage projects.
“Recent innovations in AI, with techniques such as FNOs, can accelerate calculations by orders of magnitude, taking an important step in helping scale carbon capture and storage technologies,” said Ranveer Chandra, chief executive of the company. industry research at Microsoft and contributor to the Northern Lights Initiative, a large-scale carbon capture and storage project in Norway. “Our model-parallel FNO can scale up to realistic 3D problem sizes using the distributed memory of many NVIDIA Tensor Core GPUs.”
New neural operators accelerate CO2 Storage forecasts
U-FNO allows scientists to simulate how pressure levels will build up and where CO2 it will spread during the 30 years of injection. GPU acceleration with U-FNO allows you to run these 30-year simulations in a hundredth of a second on a single NVIDIA A100 Tensor Core GPU, rather than in 10 minutes using traditional methods.
With GPU-accelerated machine learning, researchers can now also quickly simulate many injection positions. Without this tool, site selection is like a shot in the dark.
The U-FNO model focuses on modeling the migration and plume pressure during the injection process, when there is the highest risk of exceeding the amount of CO2 injected. It was developed using NVIDIA A100 GPUs in the Sherlock Computing cluster at Stanford.
“For net zero to be achievable, we will need low-emission energy sources and negative-emission technologies, such as carbon capture and storage,” said Farah Hariri, U-FNO collaborator and technical manager of the projects. climate change mitigation for NVIDIA’s Earth-2, which will be the world’s first twin AI digital supercomputer. “By applying Fourier’s neural operators to carbon storage, we have shown how AI can help accelerate the climate change mitigation process. Terra-2 will exploit these techniques ”.
Learn more about U-FNO on the NVIDIA Technical Blog.
Earth-2 will use FNO-like models to address climate science challenges and contribute to global climate change mitigation efforts. Learn more about the Earth-2 and AI models used for climate science in the GTC keynote address by NVIDIA founder and CEO Jensen Huang: