A significant challenge in dimensional accuracy control of cyber-physical additive manufacturing systems (CPAMS) is the specification of geometric shape deviation models. The current practice of constructing tailor-made deviation models for each combination of computer- aided design model, additive manufacturing (AM) process, and process setting is impractical and inefficient for general application in CPAMS. We present a new framework and class of  Bayesian neural networks for automated and efficient deviation model building in CPAMS.