Speaker
Description
This work presents a novel strategy to enhance and expedite the computation of the Matrix Element Method (MEM), a powerful technique for calculating the probability of an event generated by a given theory, using generative machine learning architectures. Despite the theoretical knowledge contained in the MEM, its practical application is hindered by the need for many approximations to compute high-dimensional integrals, particularly for complex final states with jets. However, the MEM computation is desirable to obtain powerful optimal observables for EFT measurements.
Our approach employs a combination of Transformers and Normalizing Flows. The Transformer network analyzes the complete event description at the reconstruction level, extracting a latent information vector. This vector conditions a Normalizing Flow model, which learns the conditional probability at the parton level directly. The model is trained to generate plausible parton sets compatible with the observed objects, which are then used for MEM integration through importance sampling.
This strategy is scalable and can handle events with multiple jet multiplicities and additional radiation at the parton level. We will discuss the results of the initial implementation of this architecture for a complex final state, specifically the ttH(bb) semileptonic channel
