Uncertainty in atmospheric density and wind is a major cause of suboptimal
performance in the Entry, Descent, and Landing (EDL) guidance at Mars. We improve the robustness of current EDL guidance algorithms to uncertain dynamic environments by proposing a reliable on-board atmospheric estimation algorithm. The algorithm consists of a deep, recurrent neural network using an efficient architecture for time-series predictions, the Long Short-Term Memory (LSTM) cell. The LSTM network is trained on entry trajectories simulated with the Fully Numerical Predictor-corrector Guidance (FNPEG); in each trajectory the vehicle is subject to density and wind fields from instances of the Mars Global Reference Atmospheric Model (GRAM) 2010. Predictions of density and wind as a function of altitude expected along the trajectory are obtained from onboard acceleration measurements and state estimates. The algorithm achieves a RMS value over time for the relative density error in the order of 10 % for samples in the validation dataset, and significantly improves performance with respect to an exponential fit to the density.