Abstract: As satellites are progressively deployed to operate at lower orbital heights, the assumption of a constant drag for mission planning and satellite design will not hold up well. Expensive numerical simulation of satellite aerodynamic forces may be necessary to provide accurate estimations, especially at low altitudes ($<500$km). To alleviate these data requirements in building surrogate models, a physics-informed pre-training strategy is explored to embed theoretical baselines within predictions. From the assumption of free-molecular flow, residual-learning of the rarefield aerodynamics can first serve as a form of low-fidelity approximation, before sparsely learning a corrector towards the ground truth. Under data-scarce conditions, the proposed approach outperformed models trained using only data or only physics, in terms of prediction accuracy.