We present recent work in deep learning for particle physics and cosmology at NERSC, the US Dept. of Energy mission HPC centre. We will describe activity in new methods and applications; distributed training across HPC resources; and plans for accelerated hardware for deep learning in NERSC-9 (Perlmutter) and beyond.
Some of the HEP methods and applications showcased include conditional Generative Adversarial Networks on large full-detector HEP images and high-resolution dark-matter cosmology simulations; bayesian inference via probabilistic programming for LHC analyses; alternative representations of HEP data for NN training (such as GraphNNs); and architecture search approaches. We also describe computational developments and infrastructure for training these models at large scale on NERSC supercomputers through productive interfaces.