A Deep Learning based method to extract global event properties in heavy-ion collision experiments is introduced. The order invariant, point cloud representation of experimental data is chosen to train the models. The point clouds of hits recorded in detector planes or tracks reconstructed from these hits can be represented as a 2 dimensional array in which each row stores an arbitrary hit or track. The PointNet is a Deep Learning model that is designed to efficiently perform classification or regression tasks using such a representation of data. The PointNet based models are shown to accurately determine different properties of heavy-ion collisions by learning several global features of the input point cloud that can
be mapped to a target property. In particular, we demonstrate that the PointNet based models can accurately determine the collision impact parameter on event-by-event basis, and classify the events based on the nature of the QCD transition as implemented in the hydrodynamic phase of the collision. These models outperform conventional analysis techniques and can work directly on free streaming detector output with extremely high event rates.