The conventional approach to LHC analysis involves comparing the measured data to Monte Carlo simulations. These simulations start at the hard-scattering level, where the potential for new physics is maximal, and proceed through various stages, including showering, hadronization, and detector response. Unfortunately, each stage introduces complexities, resulting in a convoluted representation of the true underlying physics at the simulated detector level. Events measured at the LHC detector are also somewhat convoluted versions of the true underlying physics due to various latent effects. Eliminating these convolutions is essential for a direct comparison between theoretical predictions and measured data, which can be achieved through the process of 'Unfolding', where measured events are directly mapped to the hard-scattering level.
In this talk, I will discuss the development and application of multi-dimensional unfolding models that utilize machine-learning-based generative techniques, specifically Generative Adversarial Networks and Normalizing Flows (NFs). A key focus will be on how multi-dimensional unfolding with NFs allows the reconstruction of observables in their proper rest frame and in a probabilistically faithful way. I will highlight its practical impact through a case study on the measurement of CP-phase in the top Yukawa coupling.
