Differentiable diagnostics in video prediction models of HEDP experiments

Simulations of high-energy density experiments are functionally videos of their modeled state evolution: 2D or 3D frames of quantities such as density and temperature, visualizing the spatiotemporal dynamics of the experiment. Furthermore, because we rely on diagnostics in experiments, we must also simulate those processes via synthetic diagnostic models, often via post-processing of the state video frames. These simulations yield two complete sets of corresponding videos: one set in state space and one in diagnostic space. However, in experiments we do not know the state video, only a critically incomplete diagnostic video with missing frames, as often we can only conduct one or a few diagnostics per experiment. We have recently demonstrated that state-of-the-art video prediction machine learning models are capable of inferring that missing state sequence from the limited diagnostic sequence, but we advance these capabilities by introducing synthetic diagnostics that can be trained, in tandem, to the video prediction model by using differentiable programming. This talk focuses on the radiography and spectroscopy diagnostics, as used in the LANL radiation-wave experiment COAX and its related experiments, demonstrating that these differentiable synthetic diagnostics may bring us closer to a physically accurate inference of experimental state.