Accelerating HEDP diagnostic analysis using machine learning without surrogates

High Energy Density Physics requires robust and repeatable data analysis techniques which provide parameters and uncertainties for workhorse diagnostics such as Thomson Scattering. While gradient descent is an effective technique for solving such optimization problems, when paired with a physical model, it can be computationally cumbersome, especially when the gradients are acquired using finite differences. However, modern scientific computing libraries provide an alternative in the form of Automatic Differentiation(AD). AD enables the performant use of a physical model in conjunction with gradient descent by providing fast, exact gradients and circumvents the need for building and relying on a black-box surrogate model. In this work, Thomson scattering analysis was implemented in an AD-capable framework and showed a significant improvement in runtime by reducing the time to calculate a spectrum and the number of spectrum calculations. This technique can be applied to many inverse problems where an algorithm can effectively describe the physics of a diagnostic.