The IceCube Neutrino Observatory is a cubic-kilometer scale neutrino detector embedded in the
Antarctic ice of the South Pole. In the near future, the detector will be augmented by extensions,
such as the IceCube Upgrade and the planned Gen2 detector. The sparseness of observed light
in the detector for low-energy events, and the irregular detector geometry, have always been a
challenge to the reconstruction of the detected neutrinos’ parameters of interest. This challenge
remains with the IceCube Upgrade, currently under construction, which introduces seven new
detector strings with novel detector modules. The Upgrade modules will increase the detection
rate of low-energy events and allow us to further constrain neutrino oscillation physics. However,
the geometry of these modules render existing traditional reconstruction algorithms more difficult
to use. We introduce a new reconstruction algorithm based on Graph Neural Networks, which we
use to reconstruct neutrino events at much faster processing times than the traditional algorithms,
while providing comparable resolution. We show that our algorithm is applicable not only to
reconstructing data of the current IceCube detector, but also simulated events for next-generation
extensions, such as the IceCube Upgrade.