Graph Neural Networks in the Physical Sciences

Deep neural networks have contributed to wildly improved performance across a variety of tasks in fields like NLP and computer vision. One property of deep learning that gets credited for this progress is the ability to encoder prior knowledge of the problem structure into the architecture of the network. Unfortunately, the convenient low dimensional grid structure found in these problems that allows for straight-forward learning of locally compact filters is not present in the wide variety of non-Euclidean data used by other scientific fields which can often only be represented as graphs or point clouds. Graph Neural Networks, and in particular Graph Convolutional Neural Networks, are a family of architectures developed to address these types of problems. In this presentation, I’ll review the major ideas and milestones in the development of modern graph neural networks and highlight several recent applications in the physical sciences.