Geothermal operators face a fundamental problem: they can’t see the subsurface system they’re trying to manage. A new project aims to change that by creating a digital twin that can simulate production in a subsurface geothermal reservoir and reduce key uncertainties in geothermal production.
One goal of the Enhanced Geothermal System Twin (EGS-Twin), scheduled for release in 2029 by the US Department of Energy’s Pacific Northwest National Laboratory (PNNL), Fervo Energy, and NVIDIA, is to accelerate geothermal energy production by creating a digital replica of a real geothermal reservoir’s behavior. Once launched, the digital twin platform will be available to any geothermal plant operator to support rapid decision-making and maximize electricity generation.
Maruti Mudunuru, earth scientist and team lead at PNNL, told JPT that one of the challenges geothermal companies face is a lack of clarity around whether injection and production wells are connected through fractures.
“To characterize uncertainties, we need extensive simulation and data sets, and running multiphysics simulations takes considerable time, even on high-performance computing systems,” he said.
Geothermal companies are seeking answers to questions that will could maximize reservoir heat production, including how much fluid to inject, how much heat can be produced, which fractures the fluid is flowing through, how to better design fracture stages, and how far apart wells should be, he said. “To explore all these what-if scenarios, we need a better way to understand this in near real-time.”
Current digital twins are not scalable and usable by operators for real-time geothermal production purposes, Mudunuru said.
“What we are trying to develop is a production-level digital twin that operators can use and understand the system in near-real time. That state of the art is not there yet. This is the first step towards that state of the art,” he said.
The first priority is to firm up the understanding of fracture stages and fluid flow in the system. Combining these pieces will show how the system is behaving, he said.
The 3-year program, announced in June, is funded by DOE’s Hydrocarbons and Geothermal Energy Office.
Fervo is contributing its geothermal expertise and proprietary data from sites in Nevada and Utah. The company is also providing operational context and validation use cases. To collect subsurface data, Fervo uses fiber-optic cables and acoustic sensing technology.
NVIDIA is providing expertise and technical guidance on accelerated computing using GPUs, PhysicsNeMo, and Omniverse libraries to support development and scaling of the digital twin.
PNNL researchers will train scalable artificial intelligence (AI) models on NVIDIA AI infrastructure to process and learn from field data collected from Fervo’s EGS assets. The team will then integrate the trained AI models into the NVIDIA Omniverse libraries, enabling visualization of a physical model of the geothermal system. The project requires massive amounts of data and computing power, he said.
“A combination of this expertise from the industry and geothermal, and the AI- and ML-accelerated aspect of NVIDIA allows us to build a twin that will better understand and do things in real time because the data sets are really huge and training takes quite some computational resources,” he said.
The biggest challenges of creating a geothermal production digital twin involve handling the data and creating predictions in real time. “The streaming data is very fast” and is multimodal, Mudunuru said. “Pressure, temperature, fiber optics is one data set.”
And all that data must be connected to a conceptual model and physics simulation.
“Once you run the simulations, how do you accelerate that? Data comes at a very high frequency, like hundreds of megabytes per second, and the simulations take quite some time. So, by the time the data comes, simulations might not already be done,” he said.
Offline training on existing data paves the way. The PNNL team will train a scalable AI model of specific sites and use the resulting model as a basis to connect to NVIDIA’s tools to construct a digital twin.
Answering the Big Question
The primary question geothermal companies have is how to optimize heat production, Mudunuru said.
“I pump in cold water, I do not want to get cold water out. I want to get hot water, continuously, for longer periods of time. The primary question is: Is my fracture network from the injection well to the production well connected? How is my flow through the fractures? Is my flow going through all the fractures?,” he said.
The goal of the 3-year project is to develop a digital twin system that will enable visualization in real time of how fluid is flowing through the geothermal system to calibrate the thermal properties of the system and visualize the heat being extracted from the reservoir, Mudunuru said.
Importantly, he added, the digital twin will deliver answers about the geothermal reservoir very quickly. “I want to get an answer very quickly in a reasonable amount of time. I do not want to wait for a few weeks or months to get an answer. You can give me an answer in a few hours.”
Ultimately, the digital twin will be transferable across sites, although the ease or difficulty of this will vary depending on the similarities of the sites. “If two sites have similar geology, we can easily transfer. If two sites are of different geology, you might need to augment your database or augment a simulation database with new geological parameters,” he said.
EGS-Twin data sets like simulations and models will be anonymized so other companies can use the digital twin solution either on-premises or via the cloud for their own projects, he said. The team will test new data sets when they arrive to iterate and validate the digital twin.
