Can Computer Vision Help Relativistic Physics? \\ ML-JET dataset for relativistic heavy ion collisions

Haydar Mehryar, Chengzhi Mao, Loren Schwiebert

Keywords: Computer Vision, Pattern Recognition, CNN, PointNet, Classification, relativistic heavy ion collision, Benchmark dataset

TL;DR: Benchmark: Relativistic heavy ion collisions simulated for 3700 hours using GPUs and CPUs, generating 10.8M jet event images.

Abstract:

Understanding relativistic heavy ion collision is important to study universe evolution. Traditional methods to simulate the collision reliant on Bayesian analysis which is costly and non-scalable, and deep learning has the potential to overcome it. We present a benchmark on relativistic heavy ion collisions, which simulates the relativistic heavy ion collision for about 3700 hours on a combination of GPUs and CPUs to compute many events, producing a total of 10.8 million jet event images for benchmarking relativistic heavy ion collisions. We release it to the vision community to push forward. Our dataset converts complex physics simulations into physics images, which can be compatible with standard vision classifiers. Using the standard Convolutional Neural Networks (CNN), our initial results attain a 92% accuracy in energy loss module classification, while concurrently accelerating the simulation process by an order of magnitude and saving millions of CPU/GPU hours. Our results suggest the potential of applying computer vision algorithms to physics in particle collisions discovery and beyond.

Submission Number: 18