Post-shot Simulations with Uncertainty: Fast, Approximate Inference Using NIF Data

Post-shot simulations, where radiation hydrodynamics are tuned to match experimental data after it is collected, are foundational to the interpretation of inertial confinement fusion (ICF) and high energy density physics (HEDP) experiments. The post-shot process provides the best picture of drive parameters, stagnated plasma conditions and degradation mechanisms for ongoing experiments at the NIF and form the basis for our understanding of capsule and hohlraum designs. The standard approach, however, involves tuning a single simulation to match a small set of observables and so cannot provide uncertainty information, or allow the possibility of multiple competing explanations of the observed ICF performance.

Recent work using deep neural network surrogates and Bayesian inference has extended ‘traditional’ hand tuned postshot simulations to provide joint probability distributions over a large (~10) set of simulation input parameters. The distributions naturally capture uncertainties and allow for large sets of fundamentally different implosions which match observations equally well. While successful, this so-called “Bayesian Superpostshot” (BSPS) has proven too computationally intensive to provide timely interpretation of new experiments.

 

In this talk we will report on work to accelerate uncertain superpostshot analysis by simultaneously running rad-hydro simulations and Bayesian inference. The new approach aims to balance local and global search of the simulation input space to efficiently choose simulations that better match experimental observations. We will apply this new approach to the tuning of 1-dimensional HYDRA simulations to data collected at the NIF, and discuss the advantages over traditional postshot simulation and large-scale ensemble methods like BSPS. Finally, we will investigate the importance of a full uncertainty model and the influence it has on experimental interpretations.