Incorporating physical constraints in a deep probabilistic machine learning framework for coarse-graining dynamical systems