Utility of differentiable simulators for innovative target design in ICF/IFE

ICF/IFE target design is a complex optimisation task, requiring both predictive simulations and many input parameters. Gradient-based optimisation methods are highly efficient when there are many design variables. However, for IFE target design this gradient information needs to propagate through the simulation, which is not possible with current state-of-the-art IFE simulation codes. Automatic differentiation (AD) is a key enabling technology for machine learning applications. It allows for differentiable programming, where accurate gradient information can be computed for any computer program (roughly) automatically. In this work, we discuss the ongoing development of a 1D Lagrangian radiation-hydrodynamics code for laser direct drive simulations, lagr-ADEPT. Because it is written in JAX, it is Pythonic, GPU-native, AD-enabled, and machine learning ready. We discuss its capabilities for prediction, inverse design and uncertainty quantification for implosion targets. We examine how machine learning technologies can be coupled with this code and what this means for future high-gain design studies.