Additive manufacturing (AM) enables personalized manufacturing of low-volume products with huge varieties and geometric complexities. Control of 3D shape deformation in AM built products has been a challenging issue, particularly under a Cyber-Physical AM environment with complex data structure and data disparity. Our goal is to automate the Machine Learning (ML) of 3D shape data for fast and efficient geometric deformation control.  This talk discusses issues of ML for AM and presents our studies towards the goal, which entails prescriptive modeling of shape deformation based on limited test shapes, optimal compensation of shape deformation through a close-form solution; Bayesian learning of disparate AM data, transfer learning between different AM process conditions, and automated ML of AM data.