Speakers
Description
Unfolding, the process of correcting measured data for detector distortions, is essential for comparing collider experiments. As experimental measurements grow increasingly sophisticated, the demand for higher-dimensional unfolding methods has risen. Recently, machine learning (ML)-based unfolding approaches have emerged to address these challenges. In heavy-ion collisions, unfolding becomes particularly complex due to the presence of the underlying event. Traditionally, underlying event subtraction and unfolding are treated in separate steps. In this work, we demonstrate that ML-based unfolding methods not only outperform traditional techniques but can also integrate underlying event subtraction directly into the unfolding step, streamlining the analysis workflow.
We introduce OmniFold-HI, a heavy-ion-tailored version of the popular unfolding algorithm [1]. Our version accounts for detector acceptance, efficiency, combinatorial jets, and uncertainties, making it well-suited for real-world analyses. Furthermore, OmniFold-HI can handle an arbitrary number of observables, offering exceptional versatility. To showcase its capabilities, we apply OmniFold-HI to unfold a 15-dimensional jet substructure observable, with the presence of a high-multiplicity thermal background, and simulated detector effects. The unfolding performance is compared with traditional methods, and uncertainties are rigorously quantified. Our results are done by training and testing using distinct event generators illustrating model-independence. We also demonstrate that OmniFold-HI reproduces the maximum likelihood estimate and provide a mathematical proof of the ML-based unfolding algorithm.
[1] Andreassen et. al, Phys. Rev. Lett. 124, 182001 (2020)
| Category | Theory |
|---|
