Expansion of ATLAS at the LHC. QCD analyses beyond jets at the LHC. Final report

Jets are traditionally considered and treated as central analysis objects at hadron accelerators such as the LHC. They affect virtually every analysis, be it searches for new physics or precision measurements in the Standard Model. In particular, ATLAS and CMS have huge data sets of jets with various partonic particles on which modern analysis methods can be developed and validated. In close cooperation with the latest international developments, this project has developed the description and simulation of jets and LHC events with jets using machine learning. Four central questions were addressed within the project: 1. How can jet data be analyzed directly, without going through high-level variables, using deep neural networks? 2. Can model-independent anomaly searches be developed using modern methods from artificial intelligence using jets? 3. Can we use generative neural networks to generate LHC events with a variable number of jets from the hard process and the QCD radiation? 4. Can this forward simulation be inverted in the sense of inverse problems to unfold detector effects and jet emission? Technically speaking, all applications of machine learning in particle physics focus on control over the networks and reliable error analysis. On the one hand, theory simulations were used as a traditional research area of the group, and on the other hand, in collaboration with local experts in the field of machine learning, Bayesian neural networks, for example, were established as a central tool. All these questions were addressed in close collaboration with experimental physics groups in Heidelberg and Hamburg. The most important results are that so-called jet taggers based on deep networks will indeed play a central role in future LHC analyses, and our studies on tagging top quarks play an important role in this development. Searching for anomalies is still a promising open question, with great potential in terms of event triggers and interesting aspects of symmetries and transformer networks. Event simulations with neural networks are established as an option, as are fast detector simulations, both using modern normalizing flow networks. In particular, these networks should lead to significant progress in unfolding in a conditional setup.