Statistics of Generative Machine Learning Models in Physics

Supervisors: Jun.-Prof. Gregor Kasieczka (UHH), Jun.-Prof. Mathias Trabs (UHH)

Measurements and searches for new signatures in experimental particle physics crucially rely on high-quality simulations of the underlying particles and their interactions with complex detectors. While classical simulations provide the highest accuracy, they come with a major computational cost and their production and storage is one of the limiting factors of the physics program at the Large Hadron Collider (LHC) and future experiments.
Generative machine learning models (such as Generative adversarial models – GANs – or Variational Autoencoders VAEs) promise to speed up this generation by several orders of magnitude. The question how and if these generative models can achieve a statistical precision beyond the sample used for their training is however still largely unsolved. The successful candidate will work at the interface of machine learning, mathematical statistics, and particle physics to build accurate and fast generative models and investigate the deeper causes and conditions for successful amplification of data.