Learning Shock Hydrodynamics with Generative Models

The computational study of shock hydrodynamics involves compressible solids, liquids, and gases that undergo large deformation. These dynamic and nonlinear problems can exhibit complex instabilities. To understand some of the complex interactions that lead to these instabilities, it is possible to parameterize a hydrodynamic problem and perform a computational study using high performance computing, yielding terabytes of simulation results. Generative models can leverage complex correlations in this data to learn the mapping from simulation parameters to highly detailed full-field hydrodynamic solutions. We present our results training these models in regression to datasets involving Richtmyer-Meshkov and Rayleigh–Taylor instabilities, as well as 1D inertial confinement fusion. The learned mapping of simulation results can be used to compress, browse, and interpolate the large volume of data. When queried in real time, the generative model allows a computational scientist to quickly visualize full-field solutions within the domain, perform sensitivity analyses, and optimize complex hydrodynamic experiments.