Multidisciplinary Co-Design Optimization and Reinforcement Learning for CubeSat Architecting

Preliminary design stage of aerospace system presents a challenge to rapidly evaluate architectural selections with acceptable accuracy. Several actors involve at this stage, such as mission owner, mission designer, and technology provider. The cooperation among these actors promote co-design activities. Co-design means parts of the design are performed by "ideal" sizing models (physics-and mission-based), while the other parts leverage the flight heritage or commercial-off-the-shelf (COTS) components. However, there is barely research exploring co-design concept in both component and analysis levels using Multidisciplinary Design Optimization (MDO) technique. Therefore, we propose a framework to formulate multidisciplinary co-design optimization in these levels for CubeSat architecting. The framework is capable to quantify the impacts of integrating COTS components into an "ideally-designed" CubeSat. The results show that some COTS perform comparably as high as the ideal design, while the others do not. The performance properties (e.g. power architecture score and CubeSat mass) provide some insights to actors to decide which components are suitable to be on-board. Based on those properties, the importance of each architectural decision-making is deducted quantitatively. Then, we implement reinforcement learning (RL) into the framework to explore the design space and find the optimum CubeSat architectures. The results show that RL is advantageous compared to enumerative approach for larger number of CubeSat architectures. Both mission designer and technology provider can leverage the framework to reduce the duration of architectural decision-making during preliminary design stage.