Preview

PDF, English - main document Download (8MB) | Terms of use

Abstract

With the upcoming High-Luminosity LHC the volume of collider data will increase dramatically, leading to a new era of precision measurements. However, this also creates computational and methodological challenges. Established simulation and inference pipelines require significant upgrades to prevent them from becoming bottlenecks. This thesis investigates how generative machine learning can address these challenges. First, we investigate modern generative architectures, diffusion models and autoregressive transformers, for fast and accurate LHC event generation. We find that they can learn complex phase space distributions to percent-level precision, demonstrating their potential as surrogate simulators. Second, we advance the use of machine learning for the matrix element method, showing how generative networks can be used to encode the transfer probability and keep the phase space integration tractable. Finally, we explore high-dimensional, unbinned unfolding using generative models. We benchmark the performance of a range of methods on the same datasets and contribute several methodological advancements, including a transformer-enhanced diffusion model that achieves state-of-the-art precision.