The temperature of rock increases about 1 degree Fahrenheit for every 77 feet below the Earth’s surface. Dig far enough, and this naturally occurring hot rock, if put in contact with water, will produce steam. This steam can be converted to energy, usually in the form of electricity.
The United States is the world’s top producer of electricity from geothermal sources, and yet this electricity accounts for only 0.4 percent of the country’s energy—enough to power about 2.7 million homes annually.
Most of the nation’s geothermal energy comes from California, where sites like The Geysers, the Imperial Valley Geothermal Project, and the Coso Geothermal Field all take advantage of existing, natural reservoirs of hot water deep underground. Engineers pipe the steam created inside the reservoirs to the surface, where the steam turns turbines and creates electricity. These sites are ideal because all the necessary ingredients for energy creation are present: hot rock, water, and steam. The problem is that most of these sites have already been exploited.
Hari Viswanathan, a hydrogeologist in the Energy and Natural Resources Security group at Los Alamos National Laboratory, says the current state of geothermal energy is not unlike the early days of the fossil fuel industry. “The industry first tapped all the most readily available sources of oil and gas, and it wasn’t until after several technological advances—like fracking and horizontal drilling—that the more difficult sites could be developed. That’s basically where the geothermal industry is at.”
In 2022, the Biden Administration announced a goal of powering at least 40 million homes with geothermal energy by 2050. But reaching that goal will require technological advancements, especially in machine learning. This is where research conducted at Los Alamos can help.
A past and present pioneer
Los Alamos has played an important role in advancing geothermal energy since the 1970s, when a ragtag bunch of scientists decided to find out if they could mine heat from the very hot, very impervious rock that’s found ubiquitously at various depths below the Earth’s surface. The challenge was reaching it (by drilling), engineering (through hydraulic fracturing) reservoirs within it, filling those reservoirs with water that heats up, bringing the hot water back to the surface, and converting it to steam to create energy. The Laboratory’s Hot Dry Rock Program, which operated just west of Los Alamos until 1995, proved all of the above was possible.
Despite the success of the Los Alamos Hot Dry Rock Program, interest in geothermal energy waned for decades. But as interest in clean energy has increased in recent years, geothermal energy is enjoying a renaissance in the form of enhanced geothermal systems (EGS)—essentially modern-day versions of the Los Alamos Hot Dry Rock Program.
To refine EGS, in 2014, the U.S. Department of Energy (DOE) funded the Frontier Observatory for Research in Geothermal Energy (FORGE). Located in Utah, FORGE operates as a testbed for the latest EGS advances. In 2021, FORGE completed construction of the world’s first angled injection well—the hole through which water is pumped down into hot rock—that will be used for EGS. The hole goes two miles deep into 430-degree-Fahrenheit granite.
“These new investments at FORGE, the flagship of our EGS research, can help us find the most innovative, cost-effective solutions and accelerate our work toward wide-scale geothermal deployment,” Secretary of Energy Jennifer Granholm said in 2022, as she announced a request for innovative research that would contribute to the site’s mission.
As geothermal drilling and fracking technology continues to improve, so has the technology that allows scientists to understand exactly what is below the Earth’s surface and how it can be most efficiently and effectively used for geothermal energy.
“The subsurface is uncertain, it’s opaque, and we have very little data on the rock in some areas,” Viswanathan says.
But now, nearly 30 years after the Hot Dry Rock Program was defunded, Los Alamos is supporting geothermal energy research once more—this time by developing machine learning technology to more accurately predict what lies beneath the Earth’s surface. Machine learning uses algorithms to parse data much faster than any human can. These algorithms can also learn, becoming better and better at their given tasks with more data and practice.
For example, humans have traditionally manually pinpointed promising geothermal sites. Experts gather known data about a region, such as type of rock, fault lines, and if silica are present (silica often indicate a geothermal reservoir). Each piece of data is assigned a value.
“Researchers find the individual attributes, or variables, in a dataset that they believe are most important to defining the presence of geothermal reservoirs,” explains Bulbul Ahmmed, a staff scientist with the Lab's Energy and Natural Resources Security group. “Then, by accounting for the importance of each variable, they compute a composite score. Based on that score, they determine if a hot spot is likely to exist.”
Los Alamos, working with DOE, has developed a machine learning program that can parse this data much more efficiently. In a test run on the Great Basin region, the Lab used machine learning to key in on 18 attributes that relate to potential geothermal hot spots. The program then produced a detailed map that indicated sites with the greatest potential. The method accurately identified sites more quickly than humans could.
However, once a potential geothermal site has been located, experts need to know as much as possible about the rock they intend to drill.
Typically, geothermal experts use seismic sensors to gather data about subsurface rock. Similar to echolocation, sensors project inaudible signals that ricochet at different speeds off of different types of rock (rock types have unique characteristics, including densities). Experts might also use magnetotelluric sensors, which infer the Earth’s subsurface features using electrical signals. Data collected from these sensors helps researchers map fault lines, natural fissures, and rock orientation (similar to the direction of wood grain). This information informs researchers about the rock’s properties, which in turn helps them understand how the rock will crack under hydraulic fracturing.
“These images of the subterranean rock that this data creates are helpful, but they leave a lot to be desired,” says Youzuo Lin, of the Lab’s Sensors and Signatures team, who notes the images are two-dimensional and may not paint a comprehensive picture of a specific area. So, to close the knowledge gap, researchers rely on something called seismic inversion, which applies physics-based mathematics to obtain the structural information of the subsurface. Another approach is to treat the lack of information as an image-to-image translation problem—basically converting an image with poor resolution to an image with higher resolution.
Lin says neither option is perfect, which is why he and his team have used machine learning to combine both methods. The new approach is called Unsupervised Physics-informed Full-Waveform Inversion (UPFWI).
“Subsurface geophysics is complicated because often we don’t have the luxury of large sets of data; there just isn’t always a lot of information for a site,” Lin says. “But aside from being able to process lots of data quickly, the advantage of machine learning is that it develops patterns, and from those patterns we can make inferences.”
UPFWI combines nearly all available information about a location: everything from field testing data to the known physics of similar rock formations. It uses forward modeling and the image-to-image method and, importantly, it then produces thousands of simulations of how the subsurface rock might appear. The might is important because these are UPFWI’s best guesses: a full range of scenarios. From there, UPFWI can produce detailed 3D images based on the average of these simulations. The end result is a more realistic depiction of the rock than any other method might produce.
“Los Alamos was the first institution to apply this type of machine learning to help characterize subsurface rock,” Lin says. “We’ve had to develop these techniques to support our national security mission [for example, to monitor for underground nuclear explosions], but they can be applied to oil and gas exploration, carbon sequestration, and certainly for mapping the subsurface characteristics of geothermal sites.”
Once operators know more or less what type of rock they’re dealing with, the next step is understanding how the rock might fracture as fluid is pumped into it, and how that fluid will then flow through the fractures. The entire system of fractures is what’s known as the reservoir.
Enter Discrete Fracture Network Modeling Suite (dfnWorks), a computer program—really a suite of machine learning algorithms—developed by Los Alamos. dfnWorks is already in use in the oil and gas and nuclear storage industries, and the technology holds great potential for EGS.
A fracked fissure is similar to a tree—a long “trunk” with “branches” radiating outward that break off into smaller “branches.” Data from seismic and other sensors can reveal larger fissures, which are sometimes meters in size. Then dfnWorks suite uses machine learning to build statistical models of the smaller fractures, which might be only millimeters in size. Each fissure, large and small, is displayed on a computer screen using an advanced spatial mesh, a way of representing mathematical data in 3D. This allows operators to view the fracked EGS reservoir almost like a 3D underground map, with the ability to rotate the image any direction desired.
“Once we’ve built the fractures on the mesh,” says Dan O’Malley, a team leader with the Lab’s Subsurface Flow and Transport group, “we can use dfnWorks to demonstrate the flow path of a fluid through the fractures, and that can help us infer how much heat the fluid will pull from the rocks.”
Because dfnWorks operates on supercomputers (as well as laptops), the program can churn out hundreds of thousands of predictive simulations within hours, making it the quickest and most accurate program of its kind.
But the Lab hasn’t stopped there.
O’Malley’s role as a quantum computing expert is to transfer dfnWork’s capabilities to the Lab’s IBM Q Network, a series of cloud-based quantum computing systems. Quantum computing uses an entirely different framework than classical computers do to solve problems—the most important difference being that quantum computing can arrive at an answer millions of times faster. O’Malley’s work, when completed, will allow Los Alamos researchers to simulate complicated EGS sites and reservoir flow patterns more accurately and quickly than anyone using classical computers.
“The holdup for EGS geothermal has really been efficiency, both in knowing how much energy we can get from a site and how to best operate the site,” O’Malley says. “With quantum computing, holds the potential to perform heroic calculations, mapping fracture networks that are massive, and doing this with more accuracy.”
In any enhanced geothermal system, a potential complication called short circuiting exists.
“In geothermal, short circuiting is a huge problem,” Viswanathan says. “Let’s say you have two wells [drilled holes deep into the rock] and a bunch of fractures in between. The ideal scenario is that you use all those fractures because it will heat your liquid more quickly. But imagine that all the water goes through one big fracture, one big artery, rather than the smaller veins. Then the hot rock will cool off and short circuit the entire site.”
To help operators constantly monitor the health and productivity of EGS sites, Los Alamos developed a machine learning program called Science-informed Machine Learning for Accelerating Real-Time Decisions in Subsurface Applications (SMART).
Funded by DOE and developed by a host of universities and national labs, including Los Alamos, SMART helps operators understand how to most efficiently run a geothermal plant, even under changing conditions. The program offers instantaneous model visualizations for the state of subsurface rock, fluid pressure management, and pressure anomaly detection that might indicate the fluid path has changed. If, over time, the fluid path does alter, SMART will not only detect changes but also offer recommendations on new operational strategies, such as changing the flow rate.
SMART allows operators to more confidently make decisions and get the most out of current and future geothermal plants. “As the United States begins to develop geothermal sites that are lower in temperature, and there are many of these across the nation, knowing how to effectively extract heat from these sites becomes very important,” Viswanathan says. “It could be a big game changer.”
The subsurface can be deceptively delicate, but under the Earth’s surface, fault lines, tectonic forces, plate boundaries, and gravity are all interacting. Pumping fluids into that fragile system can sometimes result in things moving around—aka earthquakes.
“As you produce a subsurface reservoir, you’re changing the stress fields in these systems because of the injection and extraction of fluids,” says Paul Johnson of the Lab’s Geophysics group. “You may introduce stress on a fault that was previously inactive, so it’s important to monitor all these stresses.”
In locations such as California, Iceland, France, Switzerland, and South Korea, geothermal drilling and operation have been linked to earthquakes. In some cases, the tremors were strong enough to close the fracked fissures, thus rendering the geothermal sites obsolete. In other cases, geothermal sites have closed because local populations perceived them as dangerous.
But Johnson and his colleagues have developed machine learning tools that, using historical data as well as laboratory simulations, can recognize subtle indicators that precede an earthquake. When paired with satellite information that detects slight variations in the ground surface, this same method can also be used to monitor geothermal sites during drilling and operation so experts can recognize early earthquake warning signs. With that information, they can—ideally—avoid tremors caused by geothermal construction.
“Imagine you’ve invested a tremendous amount of money into developing a geothermal location, then a governmental regulator shuts you down because you triggered an earthquake,” Johnson says. “We can help prevent this.”
Johnson sees his work as part of a bigger initiative that is critical to ensure the longevity and health of our planet. “Just imagine being able to generate electricity almost anywhere in the country from the hot rock beneath our feet,” he says. “When we think of addressing future climate change, geothermal energy has got to be one of the solutions to reaching our clean energy potential.” ★