Equivariant Neural Networks for Modeling Physical Interactions: Curve Fitting or a New Way of Understanding Nature?

Risi Kondor, University of Chicago
Computing Abstraction


In the last few years, there has been an explosion of interest in using machine learning for modeling physical and chemical systems. Research in this field ranges from using ML tools narrowly, such as to just learn the force fields that are plugged into a molecular dynamics simulation system, to trying to use AI as a drop-in replacement for PDE solvers or entire protein structure prediction pipelines. Many researchers feel that the most productive way to harness AI in science will be to tightly couple physical modeling with the more statistical, data-driven philosophy of ML. One step along this way has been the development of equivariant neural networks, which are able to explicitly account for some physical symmetries and conservation laws. In this talk, I will give a broad, somewhat mathematical, introduction to this subject and also highlight its HPC aspects.


Risi Kondor is an Associate Professor at The University of Chicago in Computer Science, Statistics, and the Computational and Applied Mathematics Initiative. Risi obtained his B.A. in Mathematics and Theoretical Physics from Cambridge, followed by an M.Sc. in Machine Learning from Carnegie Mellon and a PhD from Columbia, and postdoc positions at the Gatsby Unit (UCL) and Caltech. He has held visiting positions at Amazon Web Services, and most recently at the Flatiron Institute. Risi is interested in the confluence of machine learning, harmonic analysis, and large-scale computation. In particular, his group at the University of Chicago is pursuing a range of projects using modern neural network architectures for modeling physical systems and scientific discovery.