Inspired by passivity-based control (PBC) techniques, we propose a data-driven approach in order to learn a neural net parameterized storage function of underactuated mechanical systems. First, the PBC problem is cast as an optimization problem that searches for point estimates of the neural net parameters. Then, we improve the robustness properties of this deterministic framework against system parameter uncertainties and measurement error by injecting techniques from Bayesian inference. In the Bayesian framework, the neural net parameters are samples drawn from a posterior distribution learned via Variational Inference. We demonstrate the performance of the deterministic and Bayesian trainings on the swing-up task of an inertia wheel pendulum in simulation and realworld experiments.