Spacecraft Component and Configuration Design with Reinforcement Learning Trained Transformers

Spacecraft design, including both component design and configuration design, is a complex, multi-domain problem with a large design space. Thorough exploration of the design space is important for minimizing cost while achieving performance objectives, as improvements in the early stage of design can have large downstream effects. This process relies heavily on the experience and intuition of engineers which limits the exploration of the design space. Given this, a method which will help teams of engineers explore the design space of spacecraft designs more thoroughly is desirable. For this problem both spacecraft component design and spacecraft configuration design are considered simultaneously instead of separately as they are typically treated both in the literature and in practice. To solve the problem, a transformer is trained using reinforcement learning. This model is then compared to a genetic algorithm and random search exploring the joint spacecraft component and configuration design space. Results indicate that while the reinforcement learning-trained transformer successfully learns to generate valid designs, it struggles to optimize on the objectives within the space of valid designs when compared to a genetic algorithm. This is in contrast to results found in the configuration-only problem, and further exploration into the capabilities and limitations of the method is necessary.