Many countries have pledged to achieve net zero emissions by 2050. But making that goal a reality will require complicated feats of engineering involving the multilayered physical system that is planet Earth, says Jef Caers, professor of geological sciences at the Stanford Doerr School of Sustainability and a faculty member of the Stanford Institute for Human-Centered AI. It will also require speedy and precise planning to decide what should be done when, where, and in what order.
“Doing this successfully will require AI,” Caers says. “We can’t figure it out on our own.”
While environmental AI approaches vary, Caers is mostly interested in using AI for planning and decision making. His team uses a type of statistical reasoning called “sequential planning under uncertainty.” Think of it as a huge, complicated chess game. Each move affects the ensuing move, and it’s important to make the best choices possible along the way given the information available at each point in time.
In an early proof of concept, Caers and his colleagues collaborated with the company KoBold Metals to develop an AI system for quickly and efficiently deciding where to drill for minerals that are essential for making electric vehicle batteries. “It’s a way of prospecting in an intelligent way by deciding what information needs to be gathered in order to reduce uncertainty most,” Caers says.
The team recently published the work under the name Intelligent Prospector, and it can be readily rejiggered to address additional net zero challenges, Caers says.
For example, Caers’ team has started using the same approach to determine where and how to inject carbon dioxide (CO2) into the earth to safely store CO2 produced by steel or cement factories. A prototype they developed with HAI funding showed that by using AI to make smart injection decisions, it’s possible to increase the amount of CO2 stored at a given site, relative to conventional optimization approaches. That work has now prompted a $5.15 million collaboration with the Austrian energy company OMV and professor Mykel Kochenderfer (Stanford Engineering) that Caers’ team will use to develop intelligent agents to help Europe transition from oil to sustainable energy sources for residential and industrial heat.
“The software we use for all of these projects is very similar,” Caers says. “Whether you’re planning to drill wells for geothermal energy or to monitor a CO2 injection site over time, there’s a single formulation of this problem mathematically.” This means, he says, that there will be many other instances in which AI can contribute to planning for Net Zero 2050.
Net Zero 2050 and CO2 Sequestration
To reach net zero emissions by 2050, countries will have to enact numerous subgoals such as transitioning to electric vehicles, improving battery technology, and increasing the use of energy sources such as wind, solar, and geothermal. But perhaps the toughest challenge will be dealing with industrial heat production, Caers says. High-quality heat of 400 C to more than 1,400 C is required to make certain essential materials such as steel, cement, glass, and plastics. “That’s a big problem because fossil fuels burn hot and today are relatively cheap,” Caers says. In the long run, other solutions are emerging, such as hydrogen or ammonia, but today, the only way to mitigate the effects of burning fossil fuels for industrial heat production is to take the CO2 produced and put it underground, Caers says.
Sequestering CO2 in the earth involves injecting it into a location where porous rock lies beneath a nonporous rock such as shale that is capable of containing a substantial quantity of CO2 without leaking it into the groundwater. Knowing where and how much to inject is one planning problem Caers’ team will help to address. Another is safety: Researchers know that injecting water into the earth for geothermal energy production can trigger earthquakes. And because CO2 becomes a liquid at great depths, the same risk exists for CO2 sequestration.
The future of Net Zero 2050 depends on precise planning to ensure the risks are minimized, Caers says.
And then there are the costs. The market for carbon sequestration must be created by governments setting mandates and prices for carbon. This is already extensively happening in Europe. But even with appropriate incentives, sequestration is expensive. It requires collecting CO2 from a steel or cement plant; building pipelines to carry it to a location where it can be sequestered; injecting it into the ground; and monitoring the injection wells for leaks and seismic safety over a long period of time.
“That process has to be optimized over a timespan of 50 to 100 years, which is the complex problem we’re trying to solve with AI,” Caers says.
Prototyping an AI Decision-Making Tool
Deciding where and how fast to inject CO2 as well as monitoring for risks such as leaks and seismicity is a very complex sequential planning under uncertainty problem, Caers says. It’s made more challenging by the complexity of Earth’s rock layers, some of which contain pores where fluids can flow and others that prevent that flow. Injecting fluids into porous rock and understanding or predicting how it will bubble up is no easy feat.
To get a handle on the problem, Caers’ team recently developed a prototype system for determining how many CO2 wells to inject in what order and at what rate at a particular site. They replicated a typical situation, where, from geophysical data, they had a pretty good idea how much porous rock lay beneath a nonporous cap of dense rock. They also knew the location of seismic faults and had some information about variations in the rock’s porosity that might affect how the CO2 would move through the rock. They then simulated CO2 flowing into the rocks’ pores when injected in various places: at a random spot; at two relatively central locations chosen by an expert doing some reasonable calculations; and at a series of three sites that Caers’ team’s AI system selected to be injected in a specific order. The result: Compared with the expert’s injection site selection, the intelligent agent’s chosen procedure increased by 40% the amount of CO2 injected with no leaks. “It’s critical to come up with solutions like these that are much more optimal than a human has the capacity to imagine,” Caers says.
OMV, the Austrian energy company, was impressed. It is now collaborating with Caers’ lab to take their intelligent agent to the next step: implementation in depleted gas reservoirs in Europe. In these settings, the site has existing wells that were used to extract gas. Now the question is: Which of those locations would an intelligent agent choose for CO2 injection, and in what sequence? It’s a huge combinatorial problem – with plenty of uncertainty – that needs to be solved. “AI can solve a combinatorial problem where the number of solutions is basically infinite,” Caers says.
Investment for the Future
There is a moral imperative in achieving net zero by 2050 but companies also need a business incentive.
Investors have responded positively to the use of an intelligent agent for mining cobalt, a key mineral in electric vehicle batteries, Caers says. “They see that their return on investment is going to come faster,” Caers says. And Caers thinks investors will also be drawn to the way AI will take some of the guesswork out of carbon sequestration. The recent $5.15 million investment from OMV signals as much and means Caers and Kochernderfer’s team will double from 10 to about 20 graduate students working on the various components of the AI system.
And Caers plans to move fast.
“One of the challenges with Net Zero 2050 is the speed at which things have to happen,” Caers says. “We need to do complex planning and engineering on a massive scale at a very significant speed. And for that, we need smart solutions that use AI.”
Stanford HAI's mission is to advance AI research, education, policy, and practice to improve the human condition. Learn more.