Aerocapture is a method for orbital insertion from a hyperbolic trajectory being considered for NASA’s proposed 2030’s flagship mission to Uranus. By traveling through the planet's atmosphere to generate drag, aerocapture greatly reduces the fuel needed when firing retrograde thrusters, allowing for larger payloads or a less powerful launch vehicle. Despite these theoretical benefits, aerocapture has never flown on any planetary missions due to thin margins of error and in-situ corrections necessary to properly execute the maneuver. Critical to the guidance and control algorithms are the aerodynamic coefficients. We propose using a neural network to learn the nonlinear relationship between the raw sensor data and these aerodynamic coefficients. Specifically, we explore how network architectures designed for time dependent data, like Long-Short Term Memory (LSTM) neural networks, can produce aerodynamic coefficient estimates akin to that of a computational fluid dynamics (CFD) based lookup table, while providing more robust coefficient estimation when large environmental perturbations are experienced. Improving force and moment coefficient estimation would improve aerocapture by providing more accurate aerodynamic coefficients for use in guidance and control algorithms. This work considers multiple sensed data sources and aerodynamic coefficient data along with LSTM network architectures for model training to maximize an aerocapture maneuver’s success rate when tested in a Monte Carlo simulation. Along with this, sensitivity analyses were conducted on model hyperparameters to account for relationship complexity. Results were compared against traditional aerodynamic coefficient lookup tables within the Fully Numeric Predictor-corrector Aerocapture Guidance (FNPAG) algorithm to draw conclusions for the model’s performance.