Shocking: Scientists use artificial intelligence to improve underground carbon storage technology.

The new neural operator accelerates the simulation of carbon capture and storage, paving the way for mitigating climate change.

A team of scientists has created a brand-new artificial intelligence tool that can store greenhouse gases such as carbon dioxide in porous rocks faster and more accurately.

Carbon capture technology, also known as carbon sequestration, can transfer carbon dioxide emitted by power plants underground, thus slowing down climate change. At the same time, scientists must also avoid excessive pressure accumulation caused by injecting carbon dioxide into rocks, otherwise it may destroy the geological structure and let carbon leak into aquifers and even the atmosphere.

A new neural operator architecture named U-FNO can simulate the pressure level during carbon storage in milliseconds, and at the same time double the accuracy of some tasks to help scientists find the best injection speed and location. The research published in Progress of Water Resources unveiled the mystery of this operator. The co-authors of this article are from Stanford University, California Institute of Technology, Purdue University and NVIDIA.

Carbon capture and storage is one of the few methods used for decarbonization and emission reduction in oil refining, cement and steel industries. At present, more than 100 carbon capture and storage facilities are being built around the world.

U-FNO will be used to accelerate the carbon storage forecast of ExxonMobil, which has funded the research of secret operators.

James V. White, manager of underground carbon storage of ExxonMobil, said: "Reservoir simulator is an intensive computer model, which can be used by computational engineers and scientists to study multiphase flow and other complex physical phenomena in the underground geology of the earth." The machine learning technology used in this work can effectively quantify the uncertainties in large-scale underground flow models such as carbon capture and storage, and ultimately lead to better decision-making. "

How carbon storage scientists use machine learning

According to the simulation of carbon storage, scientists choose the correct injection location and speed, control the pressure accumulation, maximize the storage efficiency and ensure that the injection activities will not damage the rock stratum. Understanding the carbon dioxide plume (the diffusion of carbon dioxide underground) is also very important for the success of the storage project.

The traditional carbon sink simulator is not only time-consuming and laborious, but also has high calculation cost. The machine learning model has similar accuracy, but it can significantly reduce the required time and cost.

Based on U-Net neural network and Fourier neural operator (FNO), U-FNO can predict gas saturation and pressure accumulation more accurately. Compared with the most advanced convolutional neural network, the accuracy of U-FNO is doubled, but only one third of the training data is needed.

Anima Anandkumar, director of machine learning research at NVIDIA and professor of the Department of Computing and Mathematical Sciences at California Institute of Technology, said: "The machine learning method used for scientific modeling is completely different from the standard neural network: in the standard neural network, images with fixed resolution are generally used; In scientific modeling, images with different resolutions will be used according to the sampling method and location. The model can be summarized at different resolutions without retraining, so the speed is greatly improved. "

The trained U-FNO model can be provided through network applications, providing real-time prediction for carbon storage projects.

Ranveer Chandra, executive director of Microsoft Industry Research and partner of Norway's integrated carbon capture and storage project "Northern Lights", said: "The latest artificial intelligence innovations (such as FNO technology) can increase the computing speed by several orders of magnitude and help expand carbon capture and storage technology. An important step has been taken. At the same time, by using the distributed memory of multiple NVIDIA TensorCore GPUs, the model parallel FNO can be extended to the actual 3D problem scale. "

A new neural operator accelerates the prediction of carbon dioxide storage.

U-FNO enables scientists to simulate the pressure accumulation and diffusion position of carbon dioxide during 30 years of injection. With the GPU acceleration provided by U-FNO, scientists can simulate for 30 years in a flash only by using NVIDIA A 100 Tensor Core GPU, while it takes 10 minutes to use the traditional method.

Researchers can now use GPU to accelerate machine learning and quickly simulate multiple injection sites. Without this tool, you can only choose the place by luck.

U-FNO model focuses on simulating the migration and pressure of carbon dioxide plume during injection (the risk of over-injection of carbon dioxide is the greatest at this time). The model was developed by NVIDIA A 100 GPU of Shylock Computing Cluster of Stanford University.

Farah Hariri, collaborator of FNO University and technical director of NVIDIA Earth-2 Climate Change Mitigation Project, said: "In order to achieve zero net emission, low-emission energy and negative emission technologies, such as carbon capture and storage, are needed. The project will be the world's first AI digital twin supercomputer. By applying Fourier neural operators to carbon storage, we show how artificial intelligence can help accelerate the mitigation of climate change. Earth -2 will make full use of these technologies. "