The use of machine learning to analyze microearthquakes in the Earth’s crust is opening the possibility of a new path for the future of geothermal energy exploration, according to published research in Nature Communication from Ankur Mali, an assistant professor with USF’s Bellini College of Artificial Intelligence, Cybersecurity and Computing.
The new approach focused on the link between small seismic events and rock permeability, which is the capacity of subsurface rock to allow fluids to move. That’s a critical factor in determining whether geothermal heat can be efficiently extracted.
Rather than taking massive amounts of data and fitting an algorithm to it, the team took physics that researchers and industry professionals had already been using and trained the neural network to understand thousands of data points.
The work could have far reaching impacts on the future of a new, renewable energy resources.
Mali’s work on the research team, which was co-authored by Penn State engineering professors and graduate students, was published in Nature Communications earlier this year and recognized with the Rock Mechanics Research Award from the American Rock Mechanics Association. The association is an international engineering society that promotes collaboration among specialists, practitioners, scholars and educators in rock mechanics and geomechanics.
“There’s excitement among the team that our work is getting recognized, and that our methods can be used in diverse fields and benefit them,” Mali said. “It's not just one domain. It's benefiting multiple domains with a very wide applicability. We can potentially grasp those renewable energy sources, which are more eco-friendly, and reduce the reliability on fossil fuels.”
The research team received funding from the U.S. Department of Energy, National Science Foundation, European Research Council Advanced and European Union Next-Generation EU Grant.
Beginning with curiosity
Mali’s involvement in the research project began with a discussion with his civil engineering and geoscientist colleagues at Penn State.
At the time, the U.S. Department of Energy also had an increased interest in in geothermal energy and using that to create more sustainable energy sources.
Given Mali’s background in machine learning theory and strong foundations in mathematics and physics, his colleagues sought his perspective on formalizing and interpreting the underlying physical models.
“The problems themselves were grounded in well-understood physical principles,” he explained. “With a good foundation in physics, you can interpret the governing principles in the problem. The critical step where my expertise is required is to transition this into neural network language. Thus, learning how to approximate those equations effectively—and then encode that structure into a neural network.”
The research has been years in the making and led to several published articles as the team progresses with its findings. Several others are pending publication.
The group’s earlier published work in Nature Communications applied physics-inspired neural networks to laboratory earthquake simulations. That research laid the groundwork for applying similar techniques to real-world geothermal systems.
“The model is now physics guided,” he said. “It's not like a black box system. It is governed by physical situation. So, it follows that trajectory.”
That led to the research question: “Can we design a system that is more trustworthy and explainable, and can it be used in real world settings where we can understand the dynamics of the nature much more efficiently?”
Layering research
Mali’s work in the recent article builds on the team’s previous research.
The team’s current research not only determines how to more efficiently hydrofracture the rock by using microearthquakes as a metric, but to make that knowledge transferable and allow the same algorithm to be used worldwide.
Most deep rocks are naturally very dense, so researchers use high-pressure water to open or enlarge fractures, a process often called hydrofracturing or fracking. This method is a common method in geothermal research to improve heat extraction, and related techniques are used in some oil and gas operations.
Each pressure pulse triggered micro-earthquakes that acted like sonar pings and mapped where fractures connect. Until this research, there was limited knowledge to convert the seismic chatter into a real-time read-out of permeability.
That could increase energy production, decrease associated costs and provide a greater availability. They can also use the same system to fine tune the algorithm to avoid triggering microearthquakes.
“You don't have to drill at each location, we can analyze the seismic activity and then, based on the pattern of these rocks, you can understand where the energy is concentrated,” Mali said. “We can say, ‘OK, this is where we think the geothermal extraction point would be.’”
Mali and the team took a different approach that took physics data that researchers and industry professionals had already been using and trained the neural network to understand data.
“It's not like millions and billions of data points because now you're optimizing on the physics of the model,” Mali said. “Our dataset included a few thousand data points, which is more than sufficient to understand these things.”
That combination of physics and AI is essential for making the model both trustworthy and usable in the field.
“When you're deploying the system, or when you're trying to work on real-world problems, it's important that you trust your model,” Mali said. “That's where the physics comes into the picture by saying, ‘OK, we understand this concept. This is the physics which will help derive this. Now, use this physics to make an efficient approximation, and couple that with the neural network.’ That will give you a physics-inspired new network.”
Potential impact on renewable energy
Moving forward, testing on real-world data sets is needed, Mali said. “We need to gather more data to test how applicable or transferrable this approach is and what modification we would need at different places.”
Additional development is needed, along with much-needed funding from various governmental and private sources. Mali estimates it will still be another decade before a working prototype for harnessing the energy would be possible.
Although this research will be years in the making, it could have far reaching impact on the future of a new, renewable energy resources by accelerating the transition away from fossil fuels that we depend upon to clean energy systems that are economically competitive. The gravity of that impact isn’t lost on Mali, who keeps perspective on the project.
“I want to see the big picture of how this is going to be used, how we are going to benefit from these tools; not just developing something new, but what benefit it offers to society,” he said. “That's where this project started from, and this is where we are moving forward - one step at a time.”