GANplifying event samples

31 May 2022

Photo: SciPost (CC BY 4.0)

Visual representation of how well the true distribution is approximated by the training data (left) and generative model output (right). A value of 1.0 (white) indicates perfect description, while other values (red) indicate imperfections. The boundaries are chosen to represent areas of equal probability. One can see that the generative model describes the data more accurately over all.

How the use of generative models in particle physics can improve accuracy

Monte Carlo simulation plays an extremely important role in particle physics. However, these complex and accurate simulations take a significant amount of time. Generative machine learning models gained initial publicity for creating realistic looking "deep-fake" images of persons, but were since demonstrated to be a promising approach for speeding up physics simulations. However, one major roadblock for the practical use of generative models in physics was the question of whether the data generated by the model results in a more accurate description compared to the data used to train the model. This paper investigates this question using a series of small-scale examples and manages to show that generative models can, in fact, result in improved accuracy and are therefore a viable simulation tool.