Graph neural networks (GNNs) have become popular tools for processing physics data. A GNN is a neural network that takes as input a graph object composed of nodes, edges, and global features and outputs another graph, which could be a single global feature in the case of binary classification. Many physics datasets can be most naturally represented as a graph or a point cloud and so GNNs may be the most effective deep learning tool to analyze them. These tools can encode the geometry of our complex data without requiring a regular grid and also respect other aspects of the data structure such as permutation invariance, symmetries, variable size, etc. These tools have a range of applicability including materials discovery, clustering, image segmentation, particle tracking, etc. The goal of this tutorial is to provide a hands-on introduction to GNNs for physicists by physicists.
- Basics of Graph Neural Networks
- Graph Neural Networks for Property Prediction, atomistic optimization and material discovery
- Graph Neural Networks for Materials
- Savannah Thais, Columbia
- Ekin Dogus Cubuk, Google
- Kamal Choudhary, NIST
- Brian DeCost, NIST
Who Can Attend?
Graduate students, post-docs, and other scientists interested in learning about the exciting new area of graph neural networks. These tools are useful for point cloud data where the geometric relationships between objects is important, but they may not live on a fixed grid (as in the case of image data where convolutional neural networks have been quite successful). They are also useful in situations where the local environment is important and have found great applicability within molecular dynamics simulations as well as other materials calculations. They have also proved useful for image segmentation and prediction of polymer properties. The tutorial talks will be pedagogical and also demonstrate recent applications.