Combining Machine Learning and Physics Modeling to Predict Stimulated Brillouin Backscatter at the National Ignition Facility

ICF experiments at the NIF employ high-power lasers to drive a capsule implosion. As NIF scales up to higher laser energy, symmetry control of the implosion and preventing optics damage become increasing concerns. Stimulated Brillouin backscatter (SBS) is a laser-plasma interaction that can redirect some of the incoming laser light back along the beam path, impacting symmetry and putting optics at risk. A predictive model of SBS would address both of these concerns. Shots that are at high risk of damaging optics could be avoided, while those at low risk could forego “walk-up shots” (reduced scale shots used to test optics damage at full scale), allowing more laser time devoted to key physics experiments. Modeling of symmetry could be greatly improved, reducing the need for symmetry tuning with additional experiments.

We propose a new framework for building a machine learning model for SBS prediction. A neural network maps design parameters to potential plasma conditions inside the hohlraum. A physics simulation of CBET and SBS processes then makes a final SBS prediction, which is matched to data recorded by drive diagnostics (DrDs) across over 800 previous NIF ICF shots. The training is enabled by gradient backpropagation through the CBET and SBS physics models. We discuss the performance of the model and how to leverage its predictions in our design workflows.