Predicting performance variability of National Ignition Facility experiments

Developing truly predictive models of inertial confinement fusion experiments continues to be a challenge. High fidelity radiation hydrodynamic simulations can often do a good job at explaining the data post-shot, once uncertain input conditions like mix and drive asymmetries are tuned to match observations. However, predicting the outcome of an experiment based solely on preshot conditions requires knowing ahead of time the values of degradations an upcoming experiment is likely to experience. 

Over the last several years, we have developed techniques for inferring the distribution of simulation input conditions that are consistent with measurements for a large number of DT implosions. Recently, we’ve been using a combination of specialized sampling techniques and transfer learning with small sets of simulations to forward propagate the distribution of expected input conditions through new designs – producing predictions with quantified uncertainties.

In this talk we will present an update to this work, including our first predictions testing the extrapolation capabilities of the model from 1.9MJ to 2.2MJ laser drive experiments. We’ll also discuss how we plan to use this type of analysis to design explicitly for robust performance with exascale workflows in the coming year.

Prepared by LLNL under Contract DE-AC52-07NA27344. LLNL-ABS-865958.