Summary: Molecular dynamics (MD) simulations are a cornerstone of material science. By performing experiments in-silico, MD simulations can facilitate the screening of candidate compounds for applications in material design and drug discovery. However, the accuracy and reliability of MD simulations hinges on the choice of the potential energy function. Neural network (NN) potentials are promising due to their large model capacity. Thus, their accuracy is primarily limited by the available training data. This talk presents approaches to get the most out of the available training data and obtain trustworthy MD results with limited data. First, I will introduce relative entropy minimization as a highly data efficient training scheme that can correct numerical errors by sampling from the NN potential via an MD simulation during training. Second, I will describe the Differentiable Trajectory Reweighting method that facilitates enriching NN potentials with experimental data, in particular for larger systems inaccessible to accurate computational quantum mechanics schemes. Third, I will discuss scalable uncertainty quantification schemes that enable reliable estimation of credible intervals for MD observables. Approaches such as these pave the way towards the use of NN potential-based MD simulations in real-world decision-making.