A rapidly increasing number of satellites is orbiting Earth and collecting massive amounts of data. This data can be utilized in three principal ways: to train a machine learning (ML) model, to do inference, or to store it for later retrieval. This project is motivated by the first use case, where a ML model is trained collaboratively from distributed data within a satellite constellation.

Satellites move in deterministic orbits and have, especially in low earth orbits, very short contact times to ground stations, followed by a long offline period. This leads to unique and predictable connectivity patterns that might not conform to the underlying assumptions of most distributed ML algorithms. Moreover, depending on the communication capabilities of the satellites, i.e., whether and to what extend they have inter-satellite communication capabilities, the usage of distributed ML algorithms will either lead to tremendous delay between training iterations or requires efficient multi-hop communication techniques specifically targeted at this communication scenario.

The objective of this project is to conduct a fundamental study towards applying distributed ML in multi-hop networks with deterministic spatio-temporal structure. While primarily motivated by satellite topologies, the underlying mathematical model and developed techniques are also relevant for various kinds of mobile ad hoc and sensor networks. By exploiting the unique characteristics of this ML scenario and designing tailored algorithms, we expect to considerably reduce communication cost, energy consumption, and training time in comparison to the direct application of existing techniques. Moreover, by considering different deployment scenarios, we will gain an understanding of the impact of modifications in the system's communication capabilities.

Details