As we make the vital transition away from fossil fuels, much discussion around power has focused on where to source energy. However, Gege Wen is thinking about another, equally important concern: how to make it reliable. Renewable energy like wind or solar varies with the seasons, an unpredictability compounded by varying demands for energy. But Wen thinks that the Earth’s subsurface could be the solution. By using the subsurface as both a source of energy as well as a means to store it, we could ensure that our future is reliably powered by clean energy.
“The key question here is scale,” says Wen. “People have been injecting and producing energy from the Earth’s subsurface for a very, very long time, but whether we can adapt the subsurface industry fast enough to support the energy transition—that’s the challenge here.”
Gege Wen is a 2023 AI2050 Early Career Fellow and Assistant Professor at Imperial College London, co-appointed by the Earth Science Engineering Department and the newly launched I-X initiative on Artificial Intelligence. Her work was recognized in 2023 with the Daneshy Award for Energy Transition, and featured at the Berlin Summit for Earth Visualization Engines.
Wen develops computational methods for the earth and environmental sciences, with a focus on fulfilling society’s energy needs and transitioning to a low-carbon future. Her AI2050 project uses AI to model physics in gas storage experiments, with the ultimate goal of supporting the energy transition with subsurface carbon and energy storage. This research addresses Hard Problem #4, which involves using AI to make game-changing contributions to humanity’s greatest challenges.
The AI2050 initiative gratefully acknowledges Fayth Tan for assistance in producing this community perspective.
What is “clean firm power” and its significance to our transition away from fossil fuels?
When we combust fossil fuels, they emit CO2, [which] is what’s causing climate change. Renewable energy sources like solar and wind are clean as they emit little CO2, but they are intermittent. We can’t use them on the fly when there isn’t a natural source readily available. Clean firm power is both “clean” since it emits little or no CO2, and “firm” since it’s available on demand regardless of weather or time of day.
How would we use the Earth’s subsurface to provide or store clean firm power?
Many of the most promising clean firm power solutions rely on the Earth’s subsurface. For example, enhanced geothermal is a technology that drills into deep, hot rock, creating fractures and circulating water to extract heat and generate electricity—clean, reliable, and available 24/7. We can also retrofit existing natural gas-fired power plants with carbon capture and storage to provide a firm and flexible base load, which is the minimum amount of power needed to meet the electricity demands of a grid.
We can also store energy in the subsurface at large scale and use it like a giant battery. For example, excess electricity from wind or solar can be used to produce hydrogen via electrolysis, and then injected and stored in subsurface formations. This addresses the mismatch between intermittent renewable generation and energy demand, which is crucial to achieve a fully decarbonized, reliable power grid.
How did you first become interested in the problem of subsurface storage?
I started my work with carbon dioxide storage. At small scales, CO2 storage, geothermal, and energy storage such as hydrogen gas look very similar—they are all fluid flowing through porous media. We can study these processes by conducting flow experiments with rock cores extracted from the subsurface, by injecting gas or water into a natural rock material that has pores at different sizes and shapes.
Could you talk about how those experiments work?
It’s a very fun process! The subsurface is full of these rock structures that have a lot of heterogeneity. From a distance, it all looks similar to us. But if we take a small core from the subsurface, the pores are very different sizes. When the gas goes through these different sizes of pores, it will display completely different physics.
[Since] we found that these behaviors are heterogeneous, [we] take computed tomographic images of these rocks when we push the gas through them. It’s a CT machine, like the ones used in hospitals. We take the rock, push the fluid through it, and take high-resolution CT images that allows us to visualize how these fluids are moving. Observing how the gas actually moves through extracted rock cores will help us predict how CO2 might behave in the larger subsurface and how well-suited those subsurfaces are for storing CO2.
How do you model these properties using AI?
We use models that were originally proposed for different purposes. For example, transformers were proposed for language processing, and convolutional neural networks were proposed for visual information like pictures and videos. There is also an architecture that was specifically designed for scientific computing and modeling. The data that we are working with requires us to take features from each of these different categories and tailor them into a model that can predict our data. We’ve stitched them together into multi-modal models.
Are there any ecological consequences of carbon storage?
The consequences of carbon storage are relatively well-studied. In the US, if you’re storing carbon dioxide in the subsurface, you need to know where it’s flowing to and inform the owner of the property on the surface. If there are abandoned wells or unknown fractures or faults, CO2 could migrate along these structures and leak back to the surface. Carbon dioxide is denser than air, so it sinks and accumulates around the leakage point, making it potentially dangerous to children or small animals.
This information should be public, but a lot of it is hard to obtain and proprietary. Even when the data is published, most surface owners [cannot] make sense of it. During my Ph.D., I made a web application that allowed the general public to look at a simulation to see if it is possible that a project nearby could have led to carbon dioxide under their property. I’ve been talking with nonprofits to see if it could help communities determine whether there’s a risk.
What developments would you like to see in the field in the next few decades?
My biggest interest is in the scale. The subsurface is gigantic compared to what we see in the lab—there is a huge gap between [that] and the different scales that we can measure. [Understanding] how we can apply the learnings at this small scale to the massive subsurface is hard. It requires developments in computation and AI methodology to predict how CO2 will move at subsurface scales.
The GPT models were a big breakthrough because they are truly generalizable across language, across different lengths of text, and across different tasks. We don’t have this in physics modeling yet, but we need to get there, starting from real-world, practical problems.