TY  - GEN
Y1  - 2024///
CY  - Heidelberg
ID  - heidok34758
TI  - The Flow of LHC Events - Generative models for LHC simulations and inference
A1  - Heimel, Theo
AV  - public
N2  - Generative neural networks have various applications in LHC physics, for both fast simulations and precise inference. We first show that normalizing flows can be used to generate reconstruction-level events with percent-level precision. To estimate their generation uncertainties, we apply Bayesian neural networks. Further, we study the weight distribution from a classifier network which can be used for reweighting, as a performance metric and as a diagnostic tool. Next, we introduce the MadNIS framework for neural importance sampling. It improves classical methods for phase-space integration and sampling using adaptive multi-channel weights and normalizing flows as learnable channel mappings. We show that it leads to significant performance gains for several realistic LHC processes implemented in the MadGraph event generator. Generative networks can also improve analyses by maximizing the amount of extracted information. The matrix element method uses the full kinematic information, making it the tool of choice for small event numbers. It relies on a transfer function to model the shower, detector and acceptance effects. We show how three networks can be used to encode these effects, and for efficient phase-space integration. We use normalizing flows for fast sampling, diffusion models for precise density estimation, and solve jet combinatorics with a transformer.
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/34758/
ER  -