Automated Design of CubeSats and Small Spacecrafts

Abstract

The miniaturization of electronics, sensors and actuators has enabled the growing use of CubeSats and sub-20 kg spacecraft. Their reduced mass and volume has the potential to translate into significant reductions in required propellant and launch mass for interplanetary missions, earth observation and for astrophysics applications. There is an important need to optimize the design of these spacecraft to better ascertain their maximal capabilities by finding optimized solution, where mass, volume and power is a premium. Current spacecraft design methods require a team of experts, who use their engineering experience and judgement to develop a spacecraft design. Such an approach can miss innovative designs not thought of by a human design team. In this work we present a compelling alternative approach that extends the capabilities of a spacecraft engineering design team to search for and identify near-optimal solutions using machine learning. The approach enables automated design of a spacecraft that requires specifying quantitative goals, requiring reaching a target location or operating at a predetermined orbit for a required time. Next a virtual warehouse of components is specified that be selected to produce a candidate design. Candidate designs are produced using an artificial Darwinian approach, where fittest design survives and reproduce, while unfit individuals are culled off. Our past work in space robotic has produced systems designs and controllers that are human competitive. Finding a near-optimal solution presents vast improvements over a solution obtained through engineering judgment and point design alone. The approach shows a credible pathway to identify and evaluate many more candidate designs than it would be otherwise possible with a human design team alone.