Inertial Confinement Fusion on El Capitan

Project IceCap is combining exa-scale compute, machine learning (ML), and state of the art fusion simulations to design a robust, high yield implosion to test on the National Ignition Facility (NIF). The project has built an end-to-end iterative high performance computing (HPC) workflow that performs multi-fidelity Bayesian optimization with hohlraum and capsule simulations. IceCap combines novel optimization techniques with previously developed methods for modeling the expected variability of NIF implosions based on past experiments to ensure that we are not only optimizing for high yield, but for repeatable performance given realistic uncertainty in capsule/laser delivery. The workflow executes mini-ensembles of capsule-only simulations for candidate hohlraum designs to calculate the robustness of the design, and searches for an implosion that consistently reaches high yield despite the applied degradations. Additionally, our hohlraum simulations leverage embedded AI-accelerated physics packages, achieving significant speed-ups in the computationally expensive atomic physics calculations. In this talk, we present an overview of IceCap – the product of a focused, multi-year effort, which involved the deployment of El Capitan at Lawrence Livermore National Laboratory, adaptions of physics codes to leverage the new compute architecture, improvements to the workflow manager Merlin, and the Bayesian optimizer DeepOpt, deployment of the atomic physics neural network model Hermit, and building the CogSim variability models based on modern NIF ignition platforms.