This document summarizes key points from a workshop on machine learning for materials science held on August 1, 2019. It discusses how graphs are a natural representation for materials as they can capture local atomic environments and periodicity in crystals. A graph network framework called MEGNet is presented that achieves state-of-the-art performance for molecular and crystal property prediction. MEGNet models outperform previous methods on standard benchmarks and allow for transfer learning across properties. The workshop also covered practical considerations for training deep learning models and applications beyond bulk crystals.