AI Acceleration of HEP Collider Simulation

Abstract

Motivation• Detector simulation is a critical part of High Energy Physics (HEP), but is computationally expensive • Generative Neural Networks (GANs) have been explored as a machine learning solution (Ref.[1])• GANs can generate new simulated events from a large database of existing events, but could have mathematical and practical concerns • We use of a Convolutional Neural Network (CNN) as a solution• This CNN takes in quickly generated low-quality events and enhances them to a more useful resolution (See Fig. 1) Figure 5: Average energy in an event. Again, enhanced and high-quality match much more closely than with low-quality.Figure 2: This use of a CNN was inspired by a similar regression algorithm used by Disney Pixar to speed up the production of their computer-simulated films, where instead of enhancing images for films, we enhance photon simulation data Analysis• We want to generate a reasonable physical prediction, not predict a particular image, so pixel-by-pixel comparison may not be the only useful metric • Previous metrics for AI simulation Ref. [1] • centroid of the energy data • number of bins above a given energy threshold • We use these same metrics, but as we use a CNN and not a GAN, we can compare high quality, low quality, and CNN-enhanced events metrics against each other • Enhanced images more closely match energy centroid data (see Fig. 3 & 4) than low-quality input data • Also match aggregate energy data well (Fig. 5 & 6)