Quasi-Optimal Guidance and Control in Very Low Earth Orbit via Deep Learning for Drag-Based Collision Avoidance

The increasing number of objects in Low Earth Orbit makes active collision avoidance imperative for spacecraft operating in this region. If a propulsion system is not allowed due to safety concerns and mass or volume constraints, the collision risk can be mitigated by exploiting active aerodynamic drag modulation. When applied effectively, this approach enables the adjustment of the miss distance by appropriately varying the ballistic coefficient of one of the spacecraft, thereby reducing the likelihood of a collision. This work proposes a novel hybrid algorithm based on Deep Neural Networks and analytical relations for optimal collision avoidance in Low Earth Orbit, combining the minimization of orbital decay and the maximization of the miss distance. The training set relies on a proper pool of solutions generated with an optimal control algorithm previously developed by the authors. After discussing the results of a case study using the real-world scenario of the 3U CubeSat SOURCE, slated for launch in 2025, the algorithm has been successfully validated through a Monte Carlo analysis. A control procedure has been then identified to adapt the predicted profiles to off-nominal density and aerodynamics conditions. The introduction of a high-fidelity aerodynamic model using DSMC significantly influenced the controller design, highlighting the interdependent relationship between the aerodynamic model and the controller. This interaction underscores how each aspect impacts and shapes the other in the overall system design. The possibility of predicting the optimal control profiles with reduced error and low computational cost, even for unpredictable perturbations and high-fidelity environments, makes the proposed approach a noteworthy solution in the current space scenario.