SFU MOCAD Seminar: James Rowbottom

In this talk I will present a series of works derived from the framework of physics inspired graph neural networks (GNN). The central premise is a GNN can be seen as the discretisation of a learnable dynamical system over a graph, this allows to leverage the standard tools of numerical analysis to design and optimise in this model space. Firstly, I will demonstrate how this provides desirable architectural properties which lead to SOTA performance in common GNN node classification tasks. In the latter part of the talk, I will show how the same architectures emerge as natural candidates in a range of applications found in scientific computing including adaptive mesh refinement for finite element methods and mesh based graph inverse problems.