If you've used a large language model such as ChatGPT, you've interacted with a transformer, the neural network architecture that powers many generative artificial intelligence (AI) models.
Researchers at the U.S. Department of Energy's (DOE) Argonne National Laboratory are adapting this architecture to model complex physical systems. They're using AI transformers to accelerate and improve fluid-dynamics simulations for advanced nuclear reactors. Better simulations strengthen nuclear energy systems, keeping them safe and efficient.
In large language models, transformers analyze relationships among words and sentences. Argonne's approach applies the same underlying concept to physical data, such as locations, velocities and how fluids flow within a nuclear power plant. This helps AI find important physical relationships that affect reactor safety and performance. Predicting the effects of fluid flows is called turbulence modeling.
Researchers at Argonne are adding transformer architectures into the System Analysis Module (SAM). SAM is a tool developed at Argonne for studying advanced nuclear reactors. By combining transformer-based AI with turbulence modeling, the team aims to improve SAM's ability to show complex fluid-dynamic behavior. Traditional codes often struggle to capture these behaviors.
"With AI, we can be as accurate as the complex methods and as fast as the simple methods," said Rui Hu, principle nuclear engineer and manager of the Safety and Engineering Analysis group in Argonne's Nuclear Science and Engineering (NSE) division. "It is a union of accuracy and speed."
Traditional high-fidelity simulations can be slow. They might take minutes or hours for each calculation. On the other hand, faster, simpler approaches can be inaccurate. AI-based models can deliver results almost instantaneously while keeping the accuracy of more complex methods.
The new Argonne model in SAM has demonstrated high accuracy in representing the resistance of fluid to flow and how heat moves through it. The ability to capture these features is key to reliable turbulence modeling.
This advancement lets researchers conduct more realistic simulations. These can help improve reactor design, performance and safety. Next, the team plans to use the model to simulate entire power plants, including reactors, supporting components, and cooling, safety and auxiliary systems.
"Since we demonstrated that the architecture could show relationships between physical processes, it can be applied to other types of physics beyond fluid dynamics," said Eric Cervi, principal nuclear engineer in the NSE.
The team is also exploring the use of digital twins, or virtual models of physical systems operating in real time. They are among the first to apply transformer architectures to digital twin technology for nuclear systems. Future research will focus on expanding these models, improving their accuracy and flexibility, and integrating new AI-enabled models into SAM's simulation workflow.
By using advanced AI, Argonne is driving faster, more accurate simulations that support the safe and efficient operation of nuclear energy systems.
This work is supported by the DOE's Nuclear Energy Advanced Modeling and Simulation (NEAMS) Program.