ML-JET: A Benchmark Dataset for Classifying Heavy Ion Collisions

Understanding relativistic heavy ion collisions is important for studying the evolution of the universe. Models of these collisions require hyperparameter tuning to accurately reproduce experimental results. Traditionally, Bayesian analysis, which is costly and non-scalable, has been used for this tuning. Deep learning is being explored as a potential alternative, although it is still in the early stages. In this work, we propose a novel benchmark dataset for anomaly detection in visual data generated from relativistic heavy ion collisions using the JETSCAPE framework. This dataset consists of 10.8 million jet event images, enabling the application of computer vision techniques to this domain. Our dataset converts complex physics simulations into event images, which are 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 tuning process by an order of magni-tude and saving millions of CPU/GPU hours over Bayesian analysis. This work demonstrates the potential of leveraging AI-based anomaly detection techniques to advance both physics and computer vision, offering a scalable solution for identifying anomalies in particle collision data.