A key operation in Space Traffic Management is collision avoidance following the prediction and detection of a possible collision. With increasing space traffic, number of objects to be operated and lack of sensor maturity it is expected that ground operators will struggle and make sub-optimal decisions due to the less predictable consequences of each manoeuvre. The idea is to introduce two key ML techniques in the development of autonomous and automated space traffic management systems. More specifically, the idea is to use Machine Learning to build a global collision risk model and study the temporal evolution of the Minimum Orbit Intersection Distance (MOID) and the associated collision risk for a given space asset. In this context global means that, given a particular space asset (a known satellite) the time evolution of the MOID and associated collision risk are calculated for all possible other objects in a given orbital regime. A classifier will then trigger a collision avoidance manoeuvre after fusing the collision risk with other information (time to collision, confidence in the collision prediction, etc.). The confidence in the prediction of the collision risk is calculated by modelling the epistemic uncertainty on observations and propagating this epistemic uncertainty through the dynamics of the object. The global collision risk prediction is complemented with an online model identification algorithm based on symbolic regression. The model identification algorithm uses measurements (telemetry and tracking) to derive an analytical expression of the missing components in the dynamics of the space object in question, i.e. to reduce the epistemic uncertainty on the behaviour of the object. The model identification algorithm is then used to improve the prediction of the time evolution of the collision risk and in the derivation of a control law for the collision avoidance manoeuvre. It is proposed to use a rich GEO database for initial research development
