Data Representations for Virtual Diagnostics

High fidelity computer simulations have long been used to help

interpet experiments by providing insights in to physics that cannot

be probed with existing diagnostics. However, this process typically

involves solving challenging, high dimensional, and often

under-constraint inverse problems to find a simulation matching a

given experiment. Here, we present a different approach providing

similar insights in a simpler, faster, and more flexible manner. Using

deep learning solutions we create multi-modal data representations

from existing physics simulations that couple simulated diagnostics -

matching the experimental ones - with virtual ones, which report

unobservable properties, such as the internal plasma state of a

laser. Using manifold-constraint optimization approaches one can

subsequently project an incomplete representation of purely

experimental diagnostics onto the simulation-based data-manifold. This

implicitly provides the virtual diagnostics matching a given

experiment and thus immediate insights into otherwise inaccessible

information. Furthermore, by judiciously providing additional degrees

of freedom, virtual diagnostics are able to compensate for differences

between experiments and simulations. We will show how virtual

diagnostics enable new ways to interpret experiments and provide a

fast and convenient approach to steer experiments.