In the aerospace sector, control problems such as the steering of drones, rockets and satellites are ubiquitous. Model predictive control (MPC) has emerged as the method of choice in many applications, since it allows dealing with non-linear systems and constraints as well as the pursuit of general control goals. Repeatedly solving (generally non-convex) optimization problems online, however, renders its usage on devices with limited computational resources highly difficult or even impossible. Neural network (NN)-based controllers, on the other hand, enable inference at low cost after offline training, but oftennrequire large amounts of data for training and generalize rather poorly outside the data distribution being trained on. In recent years, the field of scientific machine learning (SciML), i.e., the combination of mechanistic and data-driven modeling, gained increasing popularity. Methods like universal differential equations (UDEs) or physics-informed neural networks (PINNs) have proven to be more (data-) efficient and generalize better than fully data-driven approaches. In this work, we present neural predictive control (NPC) as an NN-based control method that considers the dynamics of the controlled system during training. Unlike typical approaches that use imitation or reinforcement learning, we train the NN analogously to the optimization performed in MPC, backpropagating through the system’s dynamics model and therefore implicitly accounting for its physics. By also incorporating symmetries and invariances of the system into the NN architecture, we achieve control quality on par with MPC for the inverted pendulum problem, while being substantially more computation-efficient.