Adaptive Atmospheric Entry Guidance via Neural Networks

A common control strategy for atmospheric entry involves guiding the vehicle by modulating its bank angle. Bank maneuvers can be determined using predictor-corrector algorithms, where the current state is propagated to a final energy, and the distance between the target location and the propagated position is minimized by finding the optimal bank angle. Uncertainty in atmospheric density can degrade the propagation step, reducing targeting accuracy. This work addresses atmospheric uncertainty by introducing a new adaptive density model based on a recently proposed navigation solution. The navigation solution, originally designed in an open-loop fashion, estimates atmospheric density using an online-adapting neural network to correct for discrepancies between the onboard nominal model and true atmospheric density. The guidance algorithm uses this neural network, coupled with the nominal onboard model, to propagate the current state to the final target location during the prediction step of the Fully Numerical Predictor-Corrector Entry Guidance (FNPEG) framework. By incorporating this adaptive model, the accuracy of the prediction step is improved, allowing more precise guidance of the vehicle toward its target.