Forecasting the Evolution of Satellite Conjunction Event Parameters with Physics-Informed Transformers

In modern conjunction assessment workflows, operators make decisions based on collision risk estimates that change as new space object measurements arrive. This paper introduces a physics-informed transformer architecture that learns from patterns in historical conjunction event data to forecast how key parameters – including covariance matrices around object states and the probability of collision of those objects – will likely evolve with future measurements, potentially enabling more proactive risk assessment decisions even before new observations become available. The model is trained on data from three NASA satellites and an assortment of objects that closely approach those satellites, and the processes of data selection and curation, feature engineering, and model training are outlined. Preliminary results indicate that the physics-informed conjunction transformer predicts conjunction parameters at future timesteps more accurately than a naive baseline method, serving as promising evidence that making effective predictions on conjunction event evolution for better-informed maneuver decisions may be possible.