Machine learning has dramatically transformed and continues to transform how we interact with the world; however, these advances have not fully translated to the polymers domain. The reasons for this include that in polymers, we often have small datasets (due to costly experiments), sparse datasets (as the goal is often to probe specific quantities rather than a full parametrization of an entire space), stochastic materials (as polydispersity effects can be non-trivial) and the need to characterize uncertainty (to distinguish signal from noise). However, we also benefit from the existence of underlying physics. This project seeks to incorporate physical laws and domain knowledge into machine learning to improve performance with regards to small datasets, extrapolation and explainability. Approaches include, but are not limited to, transfer learning, augmentation and residual learning.