UC Santa Barbara and Lawrence Livermore National Laboratory (LLNL) have launched a three-year, $1.6 million collaboration to build machine learning models that could sharply reduce the computational cost of simulating plasmas for nuclear fusion research. The project addresses a core bottleneck in fusion development: physics simulations that are accurate but so expensive they slow down design studies and broader exploration of plasma behavior.
The work is led at UCSB by Haewon Jeong, an assistant professor of electrical and computer engineering, alongside LLNL co-primary investigator Min Sang Cho. It was one of five proposals selected from 29 applicants to LLNL's Academic Collaboration Team, with funding of $536,000 per year.
The computational problem
Nuclear fusion research depends on understanding how extreme plasmas behave when ultra-intense lasers strike solid targets, ionize material, and produce x-rays that compress fusion fuel. These plasmas exist in non-local thermodynamic equilibrium, or non-LTE states, where standard simplifying assumptions fail.
"Physics simulations can be very expensive," Jeong said. "The idea of surrogate modeling, especially AI surrogate modeling, is trying to bypass the need for solving partial differential equations, and learn from data to approximate simulation results, at a fraction of the computational cost."
Existing simulation codes can model the physics accurately, but the sheer number of simultaneous atomic interactions across time and space makes them resource-intensive. Cho said these calculations "often become a bottleneck in large-scale design studies and ensemble simulations."
What the team is building
The project, titled "Machine-learned Non-LTE Kinetics," combines neural compression, generative modeling, and time-sequence prediction. Jeong's lab has been developing these techniques across multiple scientific domains, and the fusion collaboration brings them together for a single application.
"What excites me most is that this project brings all of these growing areas of expertise together for an important scientific mission," Jeong said.
Neel Sankaran, a first-year Ph.D. student in Jeong's group, described the approach as capturing essential patterns without calculating every microscopic interaction from scratch. "The system is incredibly complex, and modeling every microscopic interaction directly is not practical," he said.
The broader field of AI for science and research has grown as labs seek faster alternatives to traditional simulation methods, though the UCSB-LLNL project focuses specifically on plasma kinetics rather than general-purpose scientific computing.
Why fusion needs faster simulations
LLNL achieved fusion ignition for the first time in 2022, a milestone Cho acknowledged while cautioning that "ignition is not the final step." Scaling fusion toward practical energy production requires deeper understanding of how radiation and energy move through fusion plasmas, particularly in non-LTE states that influence energy flow and system performance.
Faster calculations would let researchers explore a wider range of plasma conditions and understand the physical mechanisms governing radiation and energy transport. Cho said advances like this "could contribute to addressing one of the major energy challenges facing humankind."
For researchers applying machine learning to scientific problems, the project offers a concrete case study in surrogate model design. The AI learning path for research scientists covers related techniques for building predictive models in experimental and computational settings.
What success looks like
Jeong defined success in practical terms: integration into LLNL's existing simulation workflows. "If our AI surrogate model becomes integrated into those workflows and turns into a tool scientists rely on every day, that would be the ultimate success for our group," she said.
She added that the broader goal is to move AI surrogate modeling "from being an experimental research idea into a trusted part of scientific computing." Sankaran framed the work as contributing to "one of the bedrock problems for future scientific and technological development."
Why this matters for science and research professionals
The project illustrates a shift in how national laboratories approach computational bottlenecks: rather than waiting for faster hardware, teams are training ML models on existing simulation data to approximate results at lower cost. Researchers working with expensive simulations in any domain-climate, materials, fluid dynamics-can watch this collaboration as a test of whether surrogate models become accepted as daily-use tools rather than research prototypes. The integration metric Jeong named is the one to track: a model that works in a paper is common; one that gets folded into a lab's production workflow is rare.
Your membership also unlocks:
