Lawrence Livermore National Laboratory (LLNL) researchers employed an AI-driven model to predict fusion ignition days ahead of the historic 2022 shot, according to a new study in Science.
In the paper, LLNL researchers detail how they used physics-informed deep learning and a "cognitive simulation" (or CogSim) framework to forecast the success of the Dec. 5, 2022 fusion experiment at LLNL's National Ignition Facility (NIF), predicting a greater than 70% probability that it would exceed the energy breakeven point - producing more energy from the fusion reaction than the laser energy used to drive it.
"This was not a lucky guess," said Brian Spears, director for LLNL's AI Innovation Incubator and first author on the paper. "We used a rigorous, data-driven AI framework to quantify the likelihood of ignition before the shot took place, and the model, for the first time, predicted we were more likely to ignite than not. It's a new way of doing science."
Built as part of LLNL's growing CogSim toolkit - which combines AI with high-performance computing - a machine-learning (ML) model developed on LLNL's Sierra supercomputer was trained on more than 150,000 high-fidelity simulations and multiple years of experimental data from similar deuterium-tritium (DT) fusion implosions performed at NIF.
The model produced probabilistic predictions of fusion performance, complete with confidence intervals, and quantified the variability expected in fusion performance, predicting the neutron yield distribution of the upgraded 2.05 megajoule (MJ) design (N221204) that successfully demonstrated ignition. The resulting estimate showed a 74% probability of ignition - significantly higher than for previous designs - and the experimental result fell within the predicted yield range.
"In this work, we demonstrate a methodology for quantifying uncertainties associated with our most precious NIF experiments - DT high-yield attempts," said corresponding author and LLNL physicist Kelli Humbird. "This paper details our first attempt to take this information we learn by analyzing past experiments and applying it to a proposed experiment to make a prediction, with uncertainties, about the shot outcome. We shared our expectations ahead of time with inertial confinement fusion (ICF) management and were very pleased to see the experimental results fall within our estimates."
The work builds on LLNL's long-term CogSim for Science effort to fuse ML with physics-based modeling and high-performance computing. The key to scaling the CogSim approach for fusion is the use of "surrogate models" - deep neural networks that emulate LLNL's radiation hydrodynamics code HYDRA, but run orders of magnitude faster. The predictive framework works by modeling shot-to-shot variability in NIF implosions, including factors such as laser precision, capsule imperfections and asymmetries in the laser drive, and propagating those uncertainties through the surrogate HYDRA model. The result is a distribution of predicted outcomes that gives researchers a clearer view of experimental risk and reward.
"This is a powerful capability not just for predicting performance, but for guiding experimental decisions," said Spears. "It's part of our larger push toward cognitive simulation for science - where data, models and machine learning work together to support human judgment."
The team used a method called transfer learning to adapt the neural network from a previous shot design using just 57 new simulations, a fraction of what would normally be required. This allowed for rapid prediction within days, a crucial advantage for planning complex, high-stakes experiments, the researchers said.
The approach has already proven its value in subsequent fusion shots, according to the team. Repeat experiments of the same design fell within the model's predicted variability, offering a key validation of the method. The modeling capability is also now integrated into LLNL's standard fusion experiment planning workflows, providing not just a better understanding of what went right in past shots, but a data-driven lens for where to go next.
"Since [ignition] we've tested the model on many DT experiments that have been fielded at NIF," Humbird said. "We're using what we learned in this project to now explicitly design for high-yield, low-variability experiments. This will be particularly useful for shots where we might have external experiments where there is a specific range the yield needs to fall within to have a successful outcome."
Looking ahead, researchers said the model offers a new way to assess the robustness of proposed fusion designs and could help guide future experiments aimed at boosting performance. While the approach was tailored specifically for ICF, the team notes that the broader framework - combining high-fidelity simulation, experimental data and AI - may have applications in other complex systems where data is limited and experiments are expensive. Additional research would be needed to extend this methodology beyond fusion.
The research was supported with funding from the National Nuclear Security Administration. Co-authors include LLNL scientists Scott Brandon, Dan Casey, John Field, Jim Gaffney, Andrea Kritcher, Michael Kruse, Eugene Kur, Bogdan Kustowski, Steve Langer, Dave Munro, Ryan Nora, Luc Peterson, Dave Schlossberg, Paul Springer and Alex Zylstra.
