Unraveling the Universe with Petascale Graph Networks
, Princeton University
, Princeton University
Learn about using graph networks to find interpretable representations of physical laws in the universe with petabytes of data. We have 44,100 n-body simulations, each with over 20,000 nodes, where each node can be connected to ~30 other nodes. Leveraging NVIDIA’s toolkits and improving GPU utilization, we achieved over 8,000x speed-up in pre-processing. We'll demo our optimizations and graph network, which is built to back-propagate gradients through an entire simulation. One of the greatest challenges in astrophysics is understanding the relationship between galaxies and underlying parameters of the universe. Modeling such relationships with petabytes of data has been computationally prohibitive. We construct and train our graph network in an interpretable way, which allows us to contribute to existing theory by interpreting our graph network with symbolic regression in addition to providing new constraints on cosmological parameters. You don't need any particular prior knowledge for our session.