According to a new study in science, researchers at Lawrence Livermore National Laboratory (LLNL) have adopted an AI-powered model to predict fusion ignition dates prior to the historic 2022 shot.
The paper details how LLNL researchers used deep physics-based learning and “cognitive simulation” (Cogsim) frameworks, predicting the success of the fusion experiment at LLNL's National Ignition Facility (NIF) on December 5, 2022, predicting a probability of over 70%, and promoting energy beyond energy.
“This was not fortunate,” said Brian Spears, director of LLNL's AI Innovation Incubator and first author. “We used a rigorous, data-driven AI framework to quantify the likelihood of ignition before a shot occurs, and for the first time the model predicted that we are likely to fire. That's a new way of science.”
The Machine Learning (ML) model developed in LLNL's Sierra SuperComputer, built as part of the LLNL's Growth COGSIM toolkit combining AI and high performance computing, was trained on over 150,000 high fidelity simulations and multi-year experimental data with similar duterium tritium (DT) fusions performed in NIF.
This model generated stochastic predictions of fusion performance with confidence intervals, quantified the expected variability in fusion performance, and predicted neutron yield distributions, which were upgraded 2.05 Megajoule (MJ) design (N221204) with successful ignition. The results estimates showed a 74% chance of ignition (which was significantly higher than previous designs), and the experimental results fell within the predicted yield range.
“This work demonstrates a methodology for quantifying the uncertainty associated with the most valuable NIF experiments – the attempt at DT high yield.” “This paper details the first attempt learned by analyzing past experiments, applying them to the proposed experiments, and applying them to the proposed experiments to make uncertainty predictions about the shot results.
This work is based on LLNL's long-term cogsim for scientific efforts to fuse ML with physical-based modeling and high-performance computing. The key to scaling the Cogsim approach for fusion is the use of a “surrogate model.” This is a deep neural network that emulates the radiative fluid dynamics code Hydra of LLNL, but runs orders of magnitude faster. The prediction framework works by modeling shot-to-shot variations in NIF explosions, including factors such as laser accuracy, capsule defects, and laser drive asymmetry, and by propagating uncertainty via a surrogate Hydra model. The results are the distribution of predictive results that give researchers a clearer view of experimental risk and reward.
“This is a powerful ability to not only predict performance, but also guide experimental decisions,” Spears said. “This is part of our greater drive to science's cognitive simulation, where data, models and machine learning work together to support human judgment.”
The team adapted neural networks from previous shot designs using 57 new simulations using a method called transfer learning. This is part of what you normally need. This allows for quick predictions within a few days. This is an important advantage for planning complex, high-stakes experiments, the researchers said.
According to the team, this approach has already proven its value in subsequent fusion shots. Repeated experiments of the same design were within the expected variability of the model and provided important validation of the method. The modeling functionality is also integrated into LLNL's standard fusion experimental design workflow, providing a data-driven lens where you will go next, as well as a better understanding of what is right in past shots.
“Since ” [ignition] According to Humbird, he tested the model in many DT experiments fielded at the NIF. This is especially useful for shots where there may be external experiments with a specific range.
Looking ahead, the researchers said the model could provide a new way to assess the robustness of the proposed fusion design, and could help guide future experiments aimed at improving performance. Although this approach was tailored specifically for ICF, the team points out that a broader framework combining high-fidelity simulation, experimental data and AI could potentially have applications in other complex systems where data is limited and experiments are expensive. Additional research is needed to extend this method beyond fusion.
This research was supported with funding from the National Nuclear Security Agency. Co-authors include LLNL scientists Scott Brandon, Dan Casey, John Field, Jim Gaffney, Andrea Cricher, Michael Cruss, Eugene Kul, Bogdan Kustovsky, Steve Langer, Dave Munro, Ryan Nora, Lou Peterson, Dave Schrossaberg, Paul Springer and Alex Zilsla.
/Public release. This material of the Organization of Origin/Author is a point-in-time nature and may be edited for clarity, style and length. Mirage.news does not take any institutional position or aspect, and all views, positions and conclusions expressed here are the views of the authors alone.
