Machine Learning the Likelihoods

by Dr Rafał Masełek (Jožef Stefan Institute)

Thursday 26 Feb 2026, 16:00 → 17:30 Asia/Seoul

Room 1423 (KIAS)

Title: Machine Learning the Likelihoods

Speaker: Rafał Masełek

Abstract: 

Statistical models published by LHC experiments enable rigorous reinterpretation of searches for BSM Physics, but their direct use is often computationally expensive. In this talk, I present a machine-learning approach to construct fast and accurate surrogate models of profiled likelihoods derived from ATLAS analyses. Neural network regressors are trained to learn the dependence of the likelihood on variations of signal strengths across analysis bins, enabling rapid likelihood evaluation and efficient limit setting. Preliminary results for several supersymmetry searches demonstrate very good agreement with the original statistical models while achieving orders-of-magnitude reductions in computation time. This approach paves the way for scalable, efficient, and statistically faithful reinterpretation of LHC results across a broad range of BSM scenarios.