Speaker
Description
"Effective event categorization is essential for improving the sensitivity of the H → Zγ analysis. Correlations between classifier outputs and the reconstructed Zγ invariant mass can distort background distributions and complicate signal extraction. This poster presents the application of a mass-decorrelated event classifier transformer to the ATLAS Run 3 H → Zγ analysis, focusing on the Z → ℓ⁺ℓ⁻ final state, where ℓ = e or μ.
The approach combines a transformer architecture with distance correlation (DisCo) regularization to achieve signal–background discrimination while suppressing mass correlations. Studies using ATLAS Run 3 simulated samples and categories defined by the ratio of photon transverse momentum to Zγ invariant mass. Preliminary results show improved discrimination in the latter categories and smoother background mass distributions after classifier-based selection. Background-modeling studies also indicate reduced spurious-signal contributions in several categories and the applicability of simpler fit functions in some cases. These findings highlight the potential of DisCo Transformers for event categorization with better control of background mass shapes and motivate further development toward the full Run 3 analysis."
