Abstract
The advent of mega-constellations has given rise to an unprecedented exponential growth in the numbers of objects in orbit. As the number of objects sharing similar orbital trajectories increases, as do the probabilities of close encounters and subsequent collisions. Collisions produce more objects further increasing the probability of later collisions until the Earth orbit environment is rendered unusable. Accurate prediction of these encounters is key to enabling satellite operators to perform collision avoidance manoeuvres. This prediction is typically performed by a chain of 1) Orbital Propagators, to determine an objects state vector at a given time 2) Encounter Analysers, to determine which objects are sufficiently close to warrant further examination and 3) Statistical Models, to determine the probability that a conjunction will result in a collision. There is a need to rapidly compute data for a timelier response to threats as new observation data is received. The generation of both historical & augmented-future observation data can improve existing statistical models or as training data for machine learning systems. A key issue is how to ingest, propagate and provide statistics on historical observation data from CSpOC – estimated to take >20 years on a naive CPU only pipeline, and >6 months on a traditionally optimised solution.