Considerations in the Deployment of Machine Learning Algorithms on Spaceflight Hardware

Recent advances in artificial intelligence (AI) and machine learning (ML) have revolutionized many fields. ML has many potential applications in the space domain. Next generation space instruments are producing data at rates that exceed the capabilities of current spacecraft to store or transmit to ground stations. Deployment of ML algorithms onboard future spacecraft could perform processing of sensor data as it is gathered, reducing data volume and providing a dramatic increase in throughput of meaningful data. ML techniques may also be used to enhance the autonomy of space missions. However ML techniques have not yet been widely deployed in space environments, primarily due to limitations on the computational capabilities of spaceflight hardware. The need to verify that high-performance computational hardware can reliably operate in this environment delays the adoption of these technologies. Nevertheless, the availability of advanced processing capabilities onboard spacecraft is increasing. These platforms may not provide the processing power of terrestrial equivalents, but they do provide the resources necessary for deploying real-time execution of ML algorithms. In this paper, we present results exploring the implementation of ML techniques on computationally-constrained, high-reliability spacecraft hardware. We show two ML algorithms utilizing deep learning techniques which illustrate the utility of these approaches for space applications. We describe implementation considerations when tailoring these algorithms for execution on computationally-constrained hardware and present a workflow for performing these optimizations. We also present initial results on characterizing the trade space between algorithm accuracy, throughput, and reliability on a variety of hardware platforms with current and anticipated paths to spaceflight.