Knowledge extraction in Laser-plasma simulations -- A case study on why start-to-end simulations are just the beginning

Based on a recent laser-wakefield acceleratror experiment studying the electron beam dynamics during acceleration using betatron radiation diagnostics, we present the knowledge-extraction challenges in modeling recent experiments with particle-in-cell simulations such as PIConGPU. Lessons learnt: * Matching experiment and simulation results via start-to-end simulations is essential, but not the end. It is the beginning for knowledge extraction to gain physics understanding. * In-situ diagnostics toolkit needs to be flexible enough to minimize post-processing. * Reduced models help distinguishing, understanding and excluding different physics processes. * Particularly intermediate simulation states, such as particle distributions after ionization injection, need to be filterable and interfacable to other codes (--> openPMD). * Outlook: Next generation of simulations requires more than one order more data. In-situ diagnostics and machine-learning methods need to be further extended.