Inferring the complete state-vector evolution underlying HEDP experiments from limited diagnostic views

Comparing simulated HEDP results to experimental results remains challenging. Experimental data is noisy, incomplete, and invariably measures only a small portion or cross-section of a full physical state-vector. For example, radiography or spectroscopy typically yield transects that represent volume integrations over quantities such as temperature and density, the spatiotemporal evolution of such state quantities we desire. Using synthetic diagnostics to emulate the experimental measurement processes and then comparing synthetic to experimental data is a typical methodology. We propose a different approach: we train a video prediction network that learns to infer frame sequences of simulated state-vector quantities from limited synthetic diagnostics (i.e. from one synthetic radiograph predict the 10 following frames of simulated density and temperature). Once trained, we then supply experimental diagnostic data to the video prediction network to infer the corresponding evolution of state. To assess the accuracy of our inferred physical quantities, we employ a self-consistency check: specifically, we use the inferred state to generate synthetic diagnostics which are directly compared to the true experimental result. In this talk we will discuss this methodology of applying a video prediction network to form this self-consistency check and show our results on LANL HEDP experiments.